\n")
+ sys.stderr.flush()
+ sys.exit(1)
+ device, vocals_path, out_dir = sys.argv[1], sys.argv[2], sys.argv[3]
+ try:
+ _run(device, vocals_path, out_dir)
+ except Exception as e:
+ sys.stderr.write(f"@@ERROR@@{json.dumps(str(e))}\n")
+ sys.stderr.flush()
+ sys.exit(1)
+ sys.stderr.write("@@DONE@@\n")
+ sys.stderr.flush()
+
+
+if __name__ == "__main__":
+ main()
diff --git a/app/pipeline/split_runner.py b/app/pipeline/split_runner.py
new file mode 100644
index 00000000..7a8e6b95
--- /dev/null
+++ b/app/pipeline/split_runner.py
@@ -0,0 +1,154 @@
+"""Shared supervision for the on-demand vocal-split worker subprocesses.
+
+Both the lead/backing split (#275) and the duet split run the same shape of
+job: a fresh, non-persistent subprocess that loads one `audio-separator` model,
+writes a fixed set of wav files next to the job's stems, and reports over the
+stderr line protocol that `demucs_worker.py` established. Only the argv, the
+timeout and the expected output names differ, so the process lifecycle,
+stall watchdog and error hygiene live here once rather than per split.
+
+Deliberately not a persistent worker: these are occasional, user-triggered
+actions rather than the hot path every job takes, so there is no repeat
+model-load cost worth amortising (see ml-pipeline.md / #309 for why that
+reasoning applies to Demucs and not here).
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import subprocess
+import threading
+import time
+from pathlib import Path
+
+from app.core.models import Job
+from app.core.registry import set_proc
+from app.pipeline.errors import SeparationError
+
+logger = logging.getLogger("stemdeck.pipeline")
+
+_WATCHDOG_POLL_SECONDS = 30
+_TAIL_LIMIT = 40
+
+
+def _child_env() -> dict[str, str]:
+ env = os.environ.copy()
+ # Pin the child's stdio encoding to match what the parent decodes with.
+ # Without it a Windows child writes cp1252 while the parent reads utf-8, so
+ # the mismatch simply moves rather than being fixed. audio-separator emits
+ # progress bars and can echo track metadata, neither of which is
+ # guaranteed to be cp1252-safe.
+ env["PYTHONIOENCODING"] = "utf-8:replace"
+ # The worker arms a watchdog on this and hard-exits when we disappear, so a
+ # kill that runs no cleanup (SIGKILL, Force Quit, Task Manager, a crash)
+ # cannot leave it running with nobody to collect the result (#519).
+ env["STEMDECK_PARENT_PID"] = str(os.getpid())
+ try:
+ import certifi
+
+ env.setdefault("SSL_CERT_FILE", certifi.where())
+ env.setdefault("REQUESTS_CA_BUNDLE", certifi.where())
+ except ModuleNotFoundError:
+ pass
+ return env
+
+
+def run_split(
+ job: Job,
+ *,
+ argv: list[str],
+ device: str,
+ timeout: int,
+ label: str,
+ produces: list[str],
+ out_dir: Path,
+) -> list[str]:
+ """Run a split worker to completion and return `produces` on success.
+
+ Raises SeparationError (carrying the stderr tail) if the worker fails,
+ stalls past `timeout` with no output, or exits without writing every file
+ in `produces`. Never mutates anything outside `out_dir`.
+ """
+ proc = subprocess.Popen(
+ argv,
+ stdout=subprocess.DEVNULL,
+ stderr=subprocess.PIPE,
+ # utf-8/replace explicitly. text=True alone decodes with the locale
+ # encoding, which on Windows is cp1252, and a single byte outside it
+ # in a child's output kills the whole job with a UnicodeDecodeError
+ # ("'charmap' codec can't decode byte 0x8f"). This is diagnostic text:
+ # a byte we cannot read is never a reason to fail a separation.
+ text=True,
+ encoding="utf-8",
+ errors="replace",
+ bufsize=1,
+ env=_child_env(),
+ )
+ set_proc(job.id, proc)
+
+ tail: list[str] = []
+ last_output = [time.monotonic()]
+ done_evt = threading.Event()
+
+ def _watchdog() -> None:
+ while not done_evt.wait(timeout=_WATCHDOG_POLL_SECONDS):
+ if proc.poll() is not None:
+ return
+ if time.monotonic() - last_output[0] > timeout:
+ logger.warning(
+ "%s stalled for %ss with no output, terminating job %s",
+ label,
+ timeout,
+ job.id,
+ )
+ proc.terminate()
+ return
+
+ wt = threading.Thread(target=_watchdog, daemon=True)
+ wt.start()
+ job_ok = False
+ try:
+ assert proc.stderr is not None
+ for raw_line in proc.stderr:
+ last_output[0] = time.monotonic()
+ line = raw_line.strip()
+ if not line:
+ continue
+ if line == "@@DONE@@":
+ job_ok = True
+ break
+ if line.startswith("@@ERROR@@"):
+ msg = line[len("@@ERROR@@") :]
+ try:
+ msg = json.loads(msg)
+ except json.JSONDecodeError:
+ pass
+ tail.append(str(msg))
+ break
+ tail.append(line)
+ if len(tail) > _TAIL_LIMIT:
+ tail.pop(0)
+ finally:
+ done_evt.set()
+ set_proc(job.id, None)
+ wt.join(timeout=2)
+ if proc.poll() is None:
+ proc.terminate()
+ try:
+ proc.wait(timeout=5)
+ except subprocess.TimeoutExpired:
+ proc.kill()
+ proc.communicate()
+
+ if not job_ok:
+ detail = "\n".join(tail[-15:]) if tail else "(no stderr captured)"
+ logger.warning("[%s] %s failed: %s", job.id, label, detail)
+ last = tail[-1] if tail else f"{label} failed"
+ raise SeparationError(f"{label} failed: {last}", tail=tail[-_TAIL_LIMIT:], device=device)
+
+ for name in produces:
+ if not (out_dir / f"{name}.wav").is_file():
+ raise SeparationError(f"{label} did not produce {name}.wav", device=device)
+ return list(produces)
diff --git a/app/pipeline/vocal_split.py b/app/pipeline/vocal_split.py
index a3d1ebf8..c42032ff 100644
--- a/app/pipeline/vocal_split.py
+++ b/app/pipeline/vocal_split.py
@@ -8,29 +8,31 @@
occasional user-triggered action, not the hot path every job takes (the
persistent-worker pattern in ml-pipeline.md exists specifically to amortize a
cost every job pays; that reasoning doesn't apply here, #309).
+
+The subprocess lifecycle, stall watchdog and error hygiene live in
+split_runner.py, shared with the duet split.
"""
from __future__ import annotations
-import json
import logging
-import os
-import subprocess
import sys
-import threading
-import time
from pathlib import Path
from app.core.config import TIMEOUT_VOCAL_SPLIT
from app.core.models import Job
-from app.core.registry import set_proc
from app.core.settings import get_demucs_device
from app.pipeline.errors import SeparationError
+from app.pipeline.split_runner import run_split
logger = logging.getLogger("stemdeck.pipeline")
+VOCAL_SPLIT_STEMS = ["lead_vocals", "backing_vocals"]
+
def _spawn_cmd(device: str, vocals_path: Path, out_dir: Path) -> list[str]:
+ """The worker argv. Kept as its own function so tests can swap in a stub
+ worker and still exercise the real subprocess machinery."""
return [
sys.executable,
"-m",
@@ -55,102 +57,12 @@ def split_vocals(job: Job, stems_dir: Path) -> list[str]:
raise SeparationError("vocals.wav not found -- job has no vocals stem to split")
device = get_demucs_device()
- env = os.environ.copy()
- # Pin the child's stdio encoding to match what the parent now decodes with.
- # Without it a Windows child writes cp1252 while the parent reads utf-8, so
- # the mismatch simply moves rather than being fixed. Demucs and audio-
- # separator both emit progress bars and can echo track metadata, neither of
- # which is guaranteed to be cp1252-safe.
- # The worker arms a watchdog on this and hard-exits when we disappear, so a
- # kill that runs no cleanup (SIGKILL, Force Quit, Task Manager, a crash)
- # cannot leave it running with nobody to collect the result (#519).
- env["STEMDECK_PARENT_PID"] = str(os.getpid())
- env["PYTHONIOENCODING"] = "utf-8:replace"
- try:
- import certifi
-
- env.setdefault("SSL_CERT_FILE", certifi.where())
- env.setdefault("REQUESTS_CA_BUNDLE", certifi.where())
- except ModuleNotFoundError:
- pass
-
- proc = subprocess.Popen(
- _spawn_cmd(device, vocals_path, stems_dir),
- stdout=subprocess.DEVNULL,
- stderr=subprocess.PIPE,
- # utf-8/replace explicitly. text=True alone decodes with the locale
- # encoding, which on Windows is cp1252, and a single byte outside it
- # in a child's output kills the whole job with a UnicodeDecodeError
- # ("'charmap' codec can't decode byte 0x8f"). This is diagnostic text:
- # a byte we cannot read is never a reason to fail a separation.
- text=True,
- encoding="utf-8",
- errors="replace",
- bufsize=1,
- env=env,
+ return run_split(
+ job,
+ argv=_spawn_cmd(device, vocals_path, stems_dir),
+ device=device,
+ timeout=TIMEOUT_VOCAL_SPLIT,
+ label="vocal split",
+ produces=VOCAL_SPLIT_STEMS,
+ out_dir=stems_dir,
)
- set_proc(job.id, proc)
-
- tail: list[str] = []
- last_output = [time.monotonic()]
- done_evt = threading.Event()
-
- def _watchdog() -> None:
- while not done_evt.wait(timeout=30):
- if proc.poll() is not None:
- return
- if time.monotonic() - last_output[0] > TIMEOUT_VOCAL_SPLIT:
- logger.warning(
- "vocal split stalled for %ss with no output, terminating job %s",
- TIMEOUT_VOCAL_SPLIT,
- job.id,
- )
- proc.terminate()
- return
-
- wt = threading.Thread(target=_watchdog, daemon=True)
- wt.start()
- job_ok = False
- try:
- assert proc.stderr is not None
- for raw_line in proc.stderr:
- last_output[0] = time.monotonic()
- line = raw_line.strip()
- if not line:
- continue
- if line == "@@DONE@@":
- job_ok = True
- break
- if line.startswith("@@ERROR@@"):
- msg = line[len("@@ERROR@@") :]
- try:
- msg = json.loads(msg)
- except json.JSONDecodeError:
- pass
- tail.append(str(msg))
- break
- tail.append(line)
- if len(tail) > 40:
- tail.pop(0)
- finally:
- done_evt.set()
- set_proc(job.id, None)
- wt.join(timeout=2)
- if proc.poll() is None:
- proc.terminate()
- try:
- proc.wait(timeout=5)
- except subprocess.TimeoutExpired:
- proc.kill()
- proc.communicate()
-
- if not job_ok:
- detail = "\n".join(tail[-15:]) if tail else "(no stderr captured)"
- logger.warning("[%s] vocal split failed: %s", job.id, detail)
- last = tail[-1] if tail else "vocal split failed"
- raise SeparationError(f"vocal split failed: {last}", tail=tail[-40:], device=device)
-
- for name in ("lead_vocals", "backing_vocals"):
- if not (stems_dir / f"{name}.wav").is_file():
- raise SeparationError(f"vocal split did not produce {name}.wav", device=device)
- return ["lead_vocals", "backing_vocals"]
diff --git a/docs/duet-training.md b/docs/duet-training.md
new file mode 100644
index 00000000..86ded67f
--- /dev/null
+++ b/docs/duet-training.md
@@ -0,0 +1,225 @@
+# Training a singer-conditioned duet separator
+
+A plan, not a record. Everything here is untried; the measured results it builds
+on are in `models.md`, and the eval harness it depends on already exists.
+
+The goal is the thing the shipped feature does not do: tell you *which* singer is
+in which stem, reliably, for a whole song.
+
+## Why conditioning, rather than a better separator
+
+The shipped model separates blindly and leaves identity to be worked out
+afterwards. That second step is where it fails, measured three ways:
+
+- output order is arbitrary per chunk -- **11 of 25 chunks** had the singers
+ swapped on a real song
+- re-assigning them afterwards with ECAPA moved speaker distance from **0.977 to
+ 0.973**, which is nothing
+- because separated audio is degraded enough that a speaker embedder cannot read
+ it reliably, which is the same reason the first two are true
+
+A conditioned model has no identification step. Give it a recording of a singer,
+get that singer back. Stem 1 is whoever was enrolled, for the entire song, by
+construction. There is no permutation to fix and nothing to infer.
+
+[IWAENC 2026](https://arxiv.org/abs/2608.14516) demonstrates the idea, taking
+target-singer SI-SDR from 0.33 to 5.58 dB. What matters is how weak their setup
+is: an Open-Unmix backbone and a **two-layer GRU with hidden size 32** as the
+singer encoder. They reached most of what this project measured (+7.3 dB) from a
+far worse starting point, which says the conditioning is carrying the result and
+nobody has yet put it on a strong backbone.
+
+Their checkpoints cannot be used: they are trained on DAMP-VSEP, which is
+research-only and reserves derivatives to Smule.
+
+## The licensing position, which is what makes this possible
+
+StemDeck is free, open source and never monetised. That is not incidental here,
+it decides the corpus.
+
+**NonCommercial is satisfied.** Every CC BY-NC corpus below is usable.
+
+**ShareAlike is satisfied by publishing the weights the same way.** Trained
+weights are data, not code. StemDeck already keeps them out of the repository and
+fetches them at first use, which is how Demucs and the models in `models.md`
+work. Publishing a checkpoint under CC BY-NC-SA next to an Apache-2.0 codebase
+changes nothing about the licence of the code.
+
+This is the difference between 20 singers and 200. Singer count is what governs
+whether a separator generalises, and it is why the VocalSet-only attempt
+recorded in `models.md` overfit.
+
+**Research-only stays out.** DAMP-VSEP reserves derivatives to Smule, and no
+amount of being open source changes that. It is the one corpus everyone else
+trains on, and skipping it is the cost of being able to ship the result.
+
+| corpus | singers | licence | note |
+|---|---|---|---|
+| OpenSinger | 93 | CC BY-NC-SA 2.0 | pop, 24 kHz |
+| JVS-MuSiC | 100 | free, check terms before use | studio, 24 kHz |
+| M4Singer | 20 | CC BY-NC-SA 4.0 | **labelled SATB**, 700 songs |
+| VocalSet | 20 | CC BY 4.0 | exercises, not songs |
+| vocadito | ~40 | CC BY 4.0 | short, real singing |
+| CSD | 1 | CC BY-NC-SA 4.0 | two keys per song |
+| GTSinger | 20 languages | CC BY-NC-SA 4.0 | 2024, multi-technique |
+
+M4Singer earns its place despite being small: it is **labelled by voice type**,
+so same-register pairs can be drawn deliberately rather than hoped for. Everywhere
+else, register has to be estimated from F0.
+
+## Architecture
+
+Keep the shipped separator and add conditioning to it.
+
+- **Backbone**: the Conv-TasNet in `app/pipeline/_convtasnet.py`, already loading
+ the CC-BY-4.0 checkpoint as a starting point. A stronger backbone is the
+ obvious follow-up, but changing two things at once makes the result
+ uninterpretable.
+- **Singer encoder**: ECAPA, which this project already runs and which is
+ Apache-2.0. It is the single largest difference from the paper above.
+- **Injection**: FiLM on the masker's bottleneck -- `gamma(z) * h + beta(z)`,
+ with `gamma` and `beta` small MLPs over the 192-dimensional embedding. FiLM
+ over concatenation because the paper found little between them and FiLM leaves
+ the backbone's shapes untouched.
+- **Output**: one stem, the enrolled singer. The second stem is the mixture minus
+ the first, which keeps mixture consistency for free and halves what the model
+ has to learn.
+
+## Data
+
+Mixtures are built, not found. Two solo recordings from **different singers of
+the same register**, summed. Ground truth is exact because the sum was
+constructed.
+
+- **Pairing**: same voice type. Use M4Singer's labels where they exist, median F0
+ elsewhere. Cross-register pairs are what the register model already handles, so
+ they are worth at most a small fraction of the batch.
+- **Enrolment**: a separate passage by the same singer, never overlapping the one
+ in the mixture. The model must key on the voice, not on the passage.
+- **Levels**: random, +/- 5 dB. Neither singer should be recoverable by being
+ the loud one.
+
+### Reverb is not the problem, and measuring it properly is why
+
+The first attempt used exponentially-decaying noise as a room. It raised the
+held-out *produced* score from +6.3 to +10.0 dB while real music fell from +6.9
+to +4.9, and the obvious reading was that the model had learned a fake room and
+that better rooms were the fix.
+
+That reading was wrong, and rebuilding the eval set with **real measured rooms**
+(OpenSLR 28, Apache-2.0: RWCP, REVERB, Aachen) showed why:
+
+| condition | synthetic rooms | real rooms |
+|---|---|---|
+| VocalSet held-out, produced | +6.55 dB | **+10.31 dB** |
+
+The synthetic reverb was far harsher than any real room. The "production gap" it
+appeared to reveal was an artefact of the augmentation, and against real rooms
+the model gives up only 1.5 dB against dry. **Reverb robustness is close to a
+non-problem, and training harder on it buys almost nothing.**
+
+Use real IRs anyway, since they are free and Apache-2.0, as a wet/dry send rather
+than convolving outright -- those IRs are 16 kHz and convolving the signal would
+teach the model that duets have no top end. But do not expect a result from it.
+
+### Measured: training on VocalSet makes real music worse
+
+Run to completion with real rooms, 8,000 steps, four singers held out:
+
+| step | produced | real music |
+|---|---|---|
+| 0 | +8.69 | **+6.93** |
+| 1500 | +9.89 | +6.06 |
+| 4000 | +9.42 | +5.52 |
+| 8000 | +9.37 | **+5.52** |
+
+Real music falls **1.4 dB and flatlines**, never recovering across the whole
+run, while the in-domain score drifts up by less than a decibel. Real rooms
+halved the damage against the synthetic-reverb attempt but did not change its
+direction.
+
+This is the clean version of the experiment: one variable, a held-out eval, and
+an out-of-domain check. **VocalSet cannot close the gap and actively hurts.** The
+stock checkpoint remains the one to ship, and the next attempt needs sung songs
+rather than better augmentation of exercises.
+
+### The gap that is real: exercises against songs
+
+What the corrected numbers leave is this:
+
+| | |
+|---|---|
+| VocalSet in a real room | +10.31 dB |
+| actual produced music | +6.93 dB |
+
+**3.4 dB, and none of it acoustic.** VocalSet is scales, arpeggios and sustained
+vowels. Real material is sung phrases with consonants, phrasing, vibrato that
+means something, doubling, compression, and two people reacting to each other.
+
+That is a training-data problem, not an augmentation one, and it is the reason
+the corpus table above matters more than anything else on this page. M4Singer is
+700 songs; OpenSinger is 50 hours of pop from 93 singers. Both are sung music
+rather than exercises, and both are reachable only because NonCommercial is
+acceptable here.
+
+## Evaluation
+
+The harness exists and is the reason the failed run was caught in five minutes
+rather than eight hours. Rebuild it from `models.md` if it has been lost:
+
+- **60 mixtures per condition**, four singers held out of training entirely
+- **two conditions**, dry and produced, because the gap between them is the
+ finding
+- **real out-of-domain audio**, produced music the model never saw
+- **95% confidence intervals**, so a sub-dB change is readable
+
+Baselines to beat, measured on the shipped model:
+
+| | |
+|---|---|
+| held-out dry | +11.78 +/- 1.38 dB |
+| held-out produced, real rooms | +10.31 +/- 1.57 dB |
+| real produced music | +6.93 +/- 1.60 dB |
+| masking ceiling | +13.3 dB |
+
+The produced figure is against real measured rooms. The +6.55 dB recorded
+earlier was against synthetic reverb and says more about that reverb than about
+the model.
+
+For a conditioned model, add the measure the whole exercise is for: **does stem 1
+hold the enrolled singer for the whole song**. Speaker distance between the
+output and the enrolment, per 30-second window, across a full track. The shipped
+model scores 0.977 there, where the true sources score 0.153.
+
+## Cost
+
+Measured on an RTX 3080, 10 GB, training only the masker (20.7M parameters):
+
+| batch | ms/step | steps/hour | VRAM |
+|---|---|---|---|
+| 16 | 130 | 27,800 | 2.9 GB |
+| 32 | 206 | 17,500 | 5.5 GB |
+| 58 | 14,200 | 254 | 9.5 GB, thrashing |
+
+Batch 16 is the working point; 58, the original recipe's batch size, does not fit
+in 10 GB. A first run of 30-60k steps is one to two hours. A serious one is days,
+not weeks.
+
+Data preparation is the part that will surprise you. The first attempt ran at 3%
+GPU because reverb used `np.convolve`: **2.6 s per step against a 130 ms step**.
+FFT convolution and a precomputed bank of rooms fixed it.
+
+## What would say this worked
+
+1. Real-world produced audio improves, **not just the produced eval set**. The
+ first attempt failed exactly here.
+2. Enrolment identity holds across a whole song: well under 0.5 speaker distance,
+ against 0.977 today.
+3. It degrades honestly on a chorus rather than returning two confident stems
+ that both contain everyone.
+
+## What this does not fix
+
+Ensembles. Extracting one enrolled singer from twelve is a harder problem than
+the two-source case, and nothing measured here suggests it comes for free. Expect
+to need more than two sources at the output and a great deal more data.
diff --git a/docs/models.md b/docs/models.md
index 317e6d27..f568bce7 100644
--- a/docs/models.md
+++ b/docs/models.md
@@ -65,3 +65,199 @@ not intend it for free redistribution at all, and its already-dead HF mirror
makes it an unreliable thing to depend on regardless of the licensing
question. `STEMDECK_KARAOKE_MODEL` remains available as an env override for a
deployment that wants to accept that risk itself.
+
+## BS-Roformer Chorus Male-Female (on-demand duet split)
+
+- **File**: `model_chorus_bs_roformer_ep_267_sdr_24.1275.ckpt` plus its
+ `config_chorus_male_female_bs_roformer.yaml`.
+- **Distributed via**: `audio-separator`, the same MIT + credit-to-UVR channel
+ as the karaoke model above. Community-trained (credited to Sucial); no
+ separate LICENSE file of its own was found, which puts it in the same
+ "unlicensed but freely and openly distributed" bucket the karaoke default
+ sits in -- not the Boosty-paywall bucket that was rejected.
+- **Credit**: Ultimate Vocal Remover (Anjok07) -- https://github.com/Anjok07/ultimatevocalremovergui
+
+`DUET_SPLIT_MODEL` in `app/core/config.py`, overridable via
+`STEMDECK_DUET_MODEL`.
+
+### What it actually separates, measured
+
+The model keys on **vocal weight/register**, not on singer identity. Measured
+on the job's own vocals stem:
+
+| material | result |
+|---|---|
+| "Shallow" (Cooper/Gaga, different registers) | clean split: 30-40 dB rejection of the idle stem through each singer's verse, per-frame energy share std 0.41 (real alternation), sum-to-mix residual 0.028 |
+| "What Is This Feeling" (Wicked, two sopranos) | no split: leads with one stem in 76% of frames and the other in 7%, which is one voice with gaps. An earlier run of the same material recorded 93/7 by energy; a fresh run gave 24/76, so the totals move and who leads is the stable signal |
+
+So this is a duet splitter for singers in different registers, and does
+nothing for same-register duets. The stem names (`voice_1` / `voice_2`)
+say that rather than promising per-singer separation.
+
+### Same-register duets (two sopranos): what was tried
+
+Splitting two singers of the *same* register -- the "Glinda and Elphaba" case --
+is the thing the register model above cannot do. What follows is the record of
+what failed, kept so the dead ends are not re-attempted, and then the model that
+works, which is MedleyVox Conv-TasNet in the next section.
+
+**What does not work.** Measured against an ideal-ratio-mask ceiling of
+**+13.4 dB SI-SDRi**, so there is real headroom.
+
+| approach | result |
+|---|---|
+| SepFormer (speech PIT separation) | two anti-correlated outputs whose sum does not reconstruct the input; out of domain on singing |
+| supervised NMF (bases per singer from their own solo passages) | -2.3 dB on held-out audio: with enough components the two dictionaries explain each other |
+| harmonic-comb masking from estimated multi-F0 | -2.7 dB. With *oracle* F0s it reaches **+8.2 dB**, so the idea is sound and estimation is what fails: multi-F0 recall 0.44-0.54, contour-to-singer assignment 0.57 (chance 0.5) |
+| unsupervised 2-way diarization of the vocals stem | fails on any song with a chorus: the dominant split is leads-vs-ensemble, not lead-vs-lead |
+| ECAPA enrolment (embed each singer from a solo passage, route windows) | routes lines well on alternating material but does not separate simultaneous singing, and did not improve a separated result: speaker distance 0.977 to 0.973. Removed. |
+| zero-shot diffusion prior (`iamycy/duet-svs-diffusion`, MIT) | **-0.91 dB SI-SDRi**, worse than not splitting, at 60x slower than realtime. The paper's 11.75 dB uses oracle retry and ground-truth conditioning |
+
+**A correction.** An earlier version of this file recorded MedleyVox as tested
+and failing -- "no split on any real duet; also splits a *solo* voice across
+both outputs". That was a mistake in how it was run, twice over. iSRNet is a
+*refinement* stage cascaded after a Conv-TasNet separator, not a separator, so
+running it alone does nothing. And the checkpoint used was from the
+`singing_librispeech` training, which separates singing from *speech*. The
+`multi_singing_librispeech` checkpoint is the one for duets and had not been
+downloaded.
+
+## MedleyVox Conv-TasNet (same-register duet split)
+
+`app/pipeline/duet_medleyvox.py`, with the model code vendored into
+`app/pipeline/_convtasnet.py`.
+
+**What it does.** Separates two singers who share a vocal range, which the
+register model above cannot do at all. Measured on ground-truth duets built from
+real production audio, against a +13.3 dB ideal-mask ceiling on the same clips:
+
+| | result |
+|---|---|
+| other singer suppressed by | **14.5 dB** (no split is 0; the ideal mask reaches 24.5) |
+| SI-SDRi | **+7.3 dB**, about 55% of the masking ceiling |
+| speed | 5-7x realtime on CPU, faster on GPU |
+| output | full input rate and channel count, stems sum back to the recording |
+
+Inference details that are worth their cost, each measured: two checkpoints of
+the same run ensembled (+0.1 to +0.2 dB), each also run on the reversed signal
+(+0.2 dB), single-pass rather than overlap-add (+0.6 dB), and masks applied to
+the original-rate spectrogram rather than resampling the output (keeps
+everything above 12 kHz, which the model itself cannot see).
+
+Things tried that did not help: shift augmentation (-0.1 dB), multi-resolution
+mask averaging (+0.04 dB), and fine-tuning on the song's own solo passages,
+which overfits within 10 steps.
+
+**Training on VocalSet made it worse.** Fine-tuning on same-register pairs from
+VocalSet (CC BY 4.0) with synthetic reverb raised the held-out *produced* score
+from +6.3 to +10.0 dB while the real-world test set fell from +6.9 to +4.9. The
+model learned the synthetic reverb rather than production. Real impulse
+responses are the open lead here.
+
+**Choosing between the two models cannot be automated.** Five blind measures
+were tried and every one scored the register model's failure as healthy; the
+detail is in `duet_medleyvox.register_split_collapsed`. The short version is
+that a wrong separation is confident and uncorrelated, exactly like a right one:
+on the failed split, stem correlation scored 0.09 against the correct output's
+0.17-0.29, and mask decisiveness scored 0.96, higher than anything else
+measured. Only the unambiguous case -- one stem essentially silent -- is
+detected automatically. `STEMDECK_DUET_SAME_REGISTER=1` forces this model for
+anyone who has listened and knows which they want.
+
+**Limits.** The checkpoints have two outputs, so an ensemble is beyond them:
+stems from two-singer passages sit at 0.32 speaker distance, and across a full
+musical number with a chorus they smear to 0.97. There is no reliable way to
+detect that from the audio in advance.
+
+### It does not work on real duets, and the eval that said otherwise was flawed
+
+Tested on "Shallow" through the running app, then checked by ear: both singers
+audible in the same stem. That report was right and the measurements agreeing
+with it were not.
+
+Speaker identity, which is the only measure that has ever caught a wrong split
+here:
+
+| | cross-stem cosine |
+|---|---|
+| clustering the **unsplit** vocals, confident windows | **0.08** |
+| clustering the unsplit vocals, all windows | 0.567 |
+| register model's stems | **0.917** |
+| same-register model's stems | 0.958 |
+
+**Naively clustering the raw vocals separates the two singers better than either
+model does.** Split by section it is no kinder: 0.952 through the solo verses,
+0.883 through the overlaps. It fails everywhere, not just where they sing at
+once.
+
+**Why the +7.3 dB and 14.5 dB figures did not predict this.** Those ground-truth
+mixtures were built by summing two *unrelated* passages -- one singer from one
+part of the song, the other from somewhere else. That is a much easier problem
+than a duet. Real duet voices are musically correlated: harmonised pitch,
+synchronised consonants, the same vowels at the same time, and a shared reverb
+and compression bus. The model separates two arbitrary voices that happen to
+overlap and does not separate two people singing together.
+
+So the SI-SDR numbers above are real but measure the wrong task. An honest eval
+needs genuine multitrack duets, where each singer's recorded track is available
+separately, rather than constructed sums.
+
+**Do not ship this as a duet splitter.** The register model's 30-40 dB figure
+recorded earlier came from rejection measured through single-singer verses,
+which is not the same claim as separating the two people.
+
+### Routing beats separating, measured
+
+Prompted by the obvious question after the failure above: if the singers can be
+told apart well enough to *decide* who is singing, why separate at all? Cut the
+vocal into windows, decide whose it is, and route it. Where both sing, say so.
+
+On "Shallow", cross-stem speaker cosine, lower being further apart:
+
+| approach | coverage | cosine |
+|---|---|---|
+| same-register model | 100% | 0.958 |
+| register model | 100% | 0.917 |
+| **routing, overlap kept in both stems** | 100% | **0.756** |
+| routing, overlap silenced | 73% | 0.626 |
+| routing, conservative | 51% | 0.467 |
+| routing, very conservative | 33% | **0.351** |
+
+**Every routing setting beats both separators**, and purity is a dial: the more
+of the song you are willing to leave out, the cleaner what remains. It is also
+far cheaper -- an ECAPA pass over the vocal, no 460 MB checkpoint, seconds
+rather than nine minutes on CPU.
+
+The reason is not subtle in hindsight. A separator has to *reconstruct* two
+signals and destroys identity doing it. Routing never touches the audio; it only
+decides where each second goes, so whatever identity was in the mixture survives
+intact.
+
+**What it still does not do**, and cannot: two people singing at once are in the
+same samples and no routing extracts them. Routing's advantage is that it can be
+honest about that, which the separators are not -- they return two confident
+stems and nothing signals that both contain everyone.
+
+**One caveat on these numbers.** At high confidence the two clusters come out
+lopsided, 2% against 31%, which suggests the clustering finds one distinctive
+voice and a remainder rather than a balanced two-singer partition. The 0.08
+"floor" quoted earlier has the same asymmetry, 92 windows against 18, so it
+flatters the ceiling. Treat the ordering above as solid and the absolute values
+as optimistic.
+
+### Licensing
+
+- **MedleyVox weights** (`Cyru5/MedleyVox`) are **CC-BY 4.0**, trained by Carson
+ Evans. Fetched at first use rather than bundled, the same pattern as Demucs.
+- **The MedleyVox inference code** (`jeonchangbin49/MedleyVox`) carries no
+ LICENSE and is **not used**. The model is four classes from `asteroid` (MIT)
+ plus one published formula, vendored into `app/pipeline/_convtasnet.py` with
+ attribution. Depending on `asteroid` directly would cost 23 transitive
+ packages including pytorch-lightning and pandas.
+- **`asteroid-filterbanks`** (MIT) is a real dependency and pulls in nothing of
+ its own.
+- **SepFormer** and the ECAPA speaker embedder (`speechbrain/*`) are Apache-2.0;
+ neither is used any more.
+- **DAMP-VSEP**, behind the singer-conditioned models in the literature, is
+ research-only and reserves derivatives to Smule. Those checkpoints are
+ unusable here regardless of quality.
diff --git a/pyproject.toml b/pyproject.toml
index ad60d5e6..161bbfcb 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -88,6 +88,12 @@ dependencies = [
# today, but pin it explicitly so this survives an eventual librosa 1.x move
# (#407) -- same platform gate as audio-separator itself.
"audioread>=3.0; sys_platform != 'darwin' or platform_machine != 'x86_64'",
+ # STFT encoder/decoder for the same-register duet split
+ # (app/pipeline/duet_medleyvox.py). MIT, and unlike the full `asteroid`
+ # package it has no dependencies of its own -- that one costs 23 transitive
+ # packages including pytorch-lightning and pandas, so the handful of model
+ # classes needed from it are vendored in app/pipeline/_convtasnet.py.
+ "asteroid-filterbanks>=0.4",
]
[project.optional-dependencies]
@@ -124,7 +130,18 @@ line-length = 100
# app/_vendor is a third-party payload copied verbatim from PyPI. Linting or
# reformatting it would create a diff against upstream that the refresh script
# would then fight every time it runs.
-exclude = ["jobs", ".run", "static/vendor", "app/_version.py", "app/_vendor"]
+# app/pipeline/_convtasnet.py is the same deal: model classes copied from
+# asteroid (MIT) so the published checkpoint keeps loading. Modernising the
+# annotations there would be harmless today and a silent trap the next time it
+# is regenerated.
+exclude = [
+ "jobs",
+ ".run",
+ "static/vendor",
+ "app/_version.py",
+ "app/_vendor",
+ "app/pipeline/_convtasnet.py",
+]
[tool.ruff.lint]
select = ["E", "F", "W", "I", "UP", "B", "SIM"]
diff --git a/static/css/daw.css b/static/css/daw.css
index 938979bb..77d3223a 100644
--- a/static/css/daw.css
+++ b/static/css/daw.css
@@ -107,7 +107,10 @@ input, textarea { font-family: inherit; }
gap: 10px;
padding: 0 14px;
flex: 1;
- min-width: 0;
+ /* A floor, not 0: below about 140px the input holds no visible text and the
+ drop-refusal message renders at zero width, which reads as no answer at
+ all. */
+ min-width: 168px;
cursor: text;
}
/* Hide decorative SVG icons in the topbar */
@@ -216,13 +219,34 @@ input, textarea { font-family: inherit; }
}
/* Stem section */
+/* Shrinks, and the URL zone does not below its floor. The composer is one
+ fixed-height row with a hard width budget, and everything the controls ask
+ for used to come out of the URL zone, which is the only flexible item in the
+ row. A third vocal mode button took the zone to nothing but its own padding
+ at 1280px: no room for the input, and none for the refusal message #584 puts
+ in the same place, so dropping an unsupported file explained itself to a
+ zero-width box. The stem chips give the space up instead, by ellipsising
+ their labels next to a colour dot that already identifies them. */
.daw-stem-section {
display: flex;
align-items: center;
gap: 10px;
padding: 0 14px;
- flex-shrink: 0;
+ flex-shrink: 1;
+ min-width: 0;
}
+.daw-stem-chips { min-width: 0; }
+.stem-choice { min-width: 0; }
+/* The dot keeps its size; only the word gives way. */
+.stem-choice > span:not(.stem-dot) {
+ overflow: hidden;
+ text-overflow: ellipsis;
+ white-space: nowrap;
+}
+/* Exempt: its buttons are a mode each, and truncating "Lead + Backing" to
+ "Lead" would name a different thing. The experimental ribbon also overhangs
+ the last button, so this must not become a clipping context. */
+.vocal-mode-toggle { flex-shrink: 0; }
.daw-extract-label {
font-size: 10px;
font-weight: 600;
@@ -303,13 +327,15 @@ input, textarea { font-family: inherit; }
color: var(--accent);
}
-/* Vocals: All / Lead + Backing (on-demand split, #275) */
+/* Vocals: All / Lead + Backing / Duets (on-demand splits, #275) */
.vocal-mode-toggle {
display: inline-flex;
height: 30px;
border: 1px solid var(--border);
border-radius: 6px;
- overflow: hidden;
+ /* Not overflow:hidden. The experimental ribbon on the last button has to
+ reach outside this box to fit its word, so the rounded ends are clipped by
+ the buttons themselves below. */
}
.vocal-mode-toggle.hidden { display: none; }
.vocal-mode-btn {
@@ -325,10 +351,67 @@ input, textarea { font-family: inherit; }
transition: background var(--t-fast), color var(--t-fast);
}
.vocal-mode-btn + .vocal-mode-btn { border-left: 1px solid var(--border); }
+.vocal-mode-btn:first-child { border-radius: 5px 0 0 5px; }
+.vocal-mode-btn:last-child { border-radius: 0 5px 5px 0; }
.vocal-mode-btn[aria-pressed="true"] {
background: color-mix(in srgb, var(--vocals) 16%, transparent);
color: var(--vocals);
}
+/* Diagonal tag marking a mode as experimental.
+
+ The text comes from data-badge, which applyTranslations fills from
+ data-i18n-badge. It cannot be a child element: the button carries data-i18n,
+ and that assigns textContent, so any child inside it is destroyed on every
+ language change.
+
+ Fitting the word is the whole difficulty. The control is 30px tall, and three
+ layouts were tried against it: a corner ribbon has a 42px chord at 45 degrees
+ and rendered "RIMEN"; running the word along the button's diagonal fitted it
+ but laid it over the label so neither could be read; letting a 45-degree
+ ribbon escape the control put it across the neighbouring stem button.
+
+ What works is a shallow angle. At 9 degrees the word needs almost no vertical
+ room, so it sits on the button's top edge, inside the button's own width, in
+ the spare height .daw-composer has above the toggle. */
+.vocal-mode-btn[data-badge] {
+ position: relative;
+}
+.vocal-mode-btn[data-badge]::after {
+ content: attr(data-badge);
+ position: absolute;
+ top: -7px;
+ left: 50%;
+ transform: translateX(-50%) rotate(-9deg);
+ padding: 1px 4px;
+ border-radius: 3px;
+ background: color-mix(in srgb, var(--vocals) 80%, transparent);
+ color: #0d1016;
+ font-size: 6px;
+ font-weight: 800;
+ line-height: 1.5;
+ letter-spacing: 0.02em;
+ text-transform: uppercase;
+ white-space: nowrap;
+ /* Sits over the button, so it must never swallow the click. */
+ pointer-events: none;
+}
+.vocal-mode-btn[data-badge][aria-pressed="true"]::after,
+.vocal-mode-btn[data-badge]:hover::after,
+.vocal-mode-btn[data-badge]:focus-visible::after {
+ background: var(--vocals);
+}
+/* Pointing at it lifts the ribbon slightly, so the warning registers before the
+ tooltip has appeared. Suppressed where motion is unwelcome. */
+.vocal-mode-btn[data-badge]:hover::after {
+ transform: translateX(-50%) rotate(-9deg) translateY(-1px) scale(1.06);
+ transition: transform var(--t-fast), background var(--t-fast);
+}
+@media (prefers-reduced-motion: reduce) {
+ .vocal-mode-btn[data-badge]:hover::after {
+ transform: translateX(-50%) rotate(-9deg);
+ transition: background var(--t-fast);
+ }
+}
/* Process button */
.daw-process-btn {
diff --git a/static/index.html b/static/index.html
index f432726d..c126cacb 100644
--- a/static/index.html
+++ b/static/index.html
@@ -83,12 +83,18 @@
Vocals
-
+
All
Lead + Backing
+ Duets
Drums
@@ -116,7 +122,7 @@
Song structure
- Experimental
+ EXPERIMENTAL
diff --git a/static/js/catalog.js b/static/js/catalog.js
index 889643b6..f890051a 100644
--- a/static/js/catalog.js
+++ b/static/js/catalog.js
@@ -721,7 +721,14 @@ async function loadTrackIntoStudio(trackId) {
// Start peaks fetch immediately — runs in parallel with job-data fetch so it
// resolves before wireUpAudio calls Multitrack.create. This prevents peaks.json
// from competing with stem WAV fetches for Safari's 6-connection-per-origin limit.
- const peaksPromise = fetch(`/api/jobs/${trackId}/stems/peaks.json`)
+ // The cache buster is not belt-and-braces. This file used to be served
+ // `immutable` for a year, which a browser takes at its word and will not
+ // revalidate even on a hard reload. The header is fixed, but any browser that
+ // cached it under the old one keeps serving that copy until the year is up,
+ // and an on-demand split rewrites this file: the new stems play under the old
+ // waveform with no way for the user to refresh out of it. A changing URL is
+ // the only thing that reaches an entry already poisoned.
+ const peaksPromise = fetch(`/api/jobs/${trackId}/stems/peaks.json?v=${Date.now()}`)
.then((r) => (r.ok ? r.json() : {}))
.catch(() => ({}));
diff --git a/static/js/constants.js b/static/js/constants.js
index 5de23c60..a9b8a311 100644
--- a/static/js/constants.js
+++ b/static/js/constants.js
@@ -10,6 +10,11 @@ export let TRACK_NAMES = ["original", ...STEM_NAMES];
// never run this split -- but the player/mixer render them as real lanes,
// swapped in for "vocals", via effectiveStemOrder() below.
export let EXTRA_STEM_NAMES = ["lead_vocals", "backing_vocals"];
+// Stems from the on-demand duet split (DUET_STEM_NAMES in
+// app/core/config.py). A separate pair, not more EXTRA_STEM_NAMES: the two
+// splits are alternatives and effectiveStemOrder() swaps "vocals" for a pair
+// only when every name in that pair is present.
+export let DUET_STEM_NAMES = ["voice_1", "voice_2"];
export async function syncStemNamesFromAPI() {
try {
@@ -23,6 +28,9 @@ export async function syncStemNamesFromAPI() {
if (Array.isArray(data.extra_stem_names) && data.extra_stem_names.length > 0) {
EXTRA_STEM_NAMES = data.extra_stem_names;
}
+ if (Array.isArray(data.duet_stem_names) && data.duet_stem_names.length > 0) {
+ DUET_STEM_NAMES = data.duet_stem_names;
+ }
} catch (e) {
console.warn("[constants] failed to sync stem names from API:", e);
}
@@ -34,8 +42,14 @@ export async function syncStemNamesFromAPI() {
// has (typically from job.stems). Order matters -- callers use this both to
// decide what to render and in what sequence (waveform stacking, mixer rows).
export function effectiveStemOrder(presentNames) {
- const splitDone = presentNames.has("lead_vocals") && presentNames.has("backing_vocals");
- return STEM_NAMES.flatMap((n) => (n === "vocals" && splitDone ? EXTRA_STEM_NAMES : [n]));
+ // Only one decomposition can be in play for a job, but check both: a job
+ // that ran the lead/backing split and was later re-run as a duet has all
+ // four on disk, and lead/backing keeps precedence so the lane order does
+ // not silently change under a user who never asked for the duet split.
+ const pair = [EXTRA_STEM_NAMES, DUET_STEM_NAMES].find((names) =>
+ names.every((n) => presentNames.has(n))
+ );
+ return STEM_NAMES.flatMap((n) => (n === "vocals" && pair ? pair : [n]));
}
// A Proxy, not a plain object, so every lookup resolves through the CURRENT
@@ -65,6 +79,8 @@ export const STEM_COLORS = {
original: "#a8b0bd",
lead_vocals: "#e8748a",
backing_vocals: "#c98fe0",
+ voice_1: "#e8748a",
+ voice_2: "#7fc7e8",
};
export const PROGRESS_COLOR = "#3a3a3a";
diff --git a/static/js/i18n.js b/static/js/i18n.js
index cae25a97..ad271fad 100644
--- a/static/js/i18n.js
+++ b/static/js/i18n.js
@@ -189,6 +189,13 @@ export function applyTranslations(root) {
scope.querySelectorAll("[data-i18n-placeholder]").forEach((el) => {
el.placeholder = t(el.getAttribute("data-i18n-placeholder"));
});
+ // Text a pseudo-element draws, via content: attr(data-badge). An element
+ // carrying data-i18n has its textContent rewritten above, so a real child
+ // element inside one is wiped on every language change; this is how such an
+ // element gets a second translated string.
+ scope.querySelectorAll("[data-i18n-badge]").forEach((el) => {
+ el.dataset.badge = t(el.getAttribute("data-i18n-badge"));
+ });
const titleKey = document.documentElement.getAttribute("data-i18n-doctitle");
if (titleKey) document.title = t(titleKey);
applyStemRowAriaLabels(scope);
@@ -233,7 +240,7 @@ const en = {
"sections.clearConfirm": "Confirm?",
"structure.toggle": "Song structure",
"structure.toggleTitle": "Experimental Song Structure Extraction: automatically label intro, verse and chorus after a split",
- "structure.experimental": "Experimental",
+ "common.experimental": "EXPERIMENTAL",
"doc.title": "StemDeck — split any track into stems",
"topbar.urlPlaceholder": "Search or drop an audio file…",
@@ -247,6 +254,8 @@ const en = {
"stem.vocals": "Vocals",
"stem.lead_vocals": "Lead Vocals",
"stem.backing_vocals": "Backing Vocals",
+ "stem.voice_1": "Voice 1",
+ "stem.voice_2": "Voice 2",
"stem.drums": "Drums",
"stem.bass": "Bass",
"stem.guitar": "Guitar",
@@ -257,6 +266,8 @@ const en = {
"vocalMode.groupAria": "Vocals mode",
"vocalMode.all": "All",
"vocalMode.split": "Lead + Backing",
+ "vocalMode.duet": "Duets",
+ "vocalMode.duetTitle": "EXPERIMENTAL. Splits the vocals into two singers. Works best when they trade lines and their voices differ. Where both sing at once, some of the other voice remains.",
"aria.mute": "Mute {name}",
"aria.solo": "Solo {name}",
@@ -379,6 +390,8 @@ const en = {
"wave.loadingAria": "Loading waveform",
"job.cancelAria": "Cancel job",
"job.cancel": "Cancel",
+ "job.splittingVocals": "Splitting lead/backing vocals…",
+ "job.splittingDuet": "Splitting duet voices…",
"footer.stemsPlaceholder": "— Stems",
"footer.stemsCount": "{count} stems",
@@ -850,7 +863,7 @@ const pl = {
"sections.clearConfirm": "Potwierdzić?",
"structure.toggle": "Struktura utworu",
"structure.toggleTitle": "Eksperymentalne wykrywanie struktury utworu: automatycznie oznacza intro, zwrotkę i refren po podziale",
- "structure.experimental": "Eksperymentalne",
+ "common.experimental": "EKSPERYMENTALNE",
"doc.title": "StemDeck — rozdziel dowolny utwór na ścieżki",
"topbar.urlPlaceholder": "Szukaj albo upuść plik audio…",
@@ -864,6 +877,8 @@ const pl = {
"stem.vocals": "Wokal",
"stem.lead_vocals": "Wokal główny",
"stem.backing_vocals": "Chórki",
+ "stem.voice_1": "Głos 1",
+ "stem.voice_2": "Głos 2",
"stem.drums": "Perkusja",
"stem.bass": "Bas",
"stem.guitar": "Gitara",
@@ -874,6 +889,8 @@ const pl = {
"vocalMode.groupAria": "Tryb wokalu",
"vocalMode.all": "Wszystko",
"vocalMode.split": "Wokal + Chórki",
+ "vocalMode.duet": "Duety",
+ "vocalMode.duetTitle": "EKSPERYMENTALNE. Dzieli wokal na dwoje śpiewaków. Działa najlepiej, gdy śpiewają na zmianę i mają różne głosy. Gdy śpiewają razem, słychać część drugiego głosu.",
"aria.mute": "Wycisz: {name}",
"aria.solo": "Solo: {name}",
@@ -996,6 +1013,8 @@ const pl = {
"wave.loadingAria": "Wczytywanie przebiegu fali",
"job.cancelAria": "Anuluj zadanie",
"job.cancel": "Anuluj",
+ "job.splittingVocals": "Dzielenie wokalu i chórków…",
+ "job.splittingDuet": "Dzielenie głosów duetu…",
"footer.stemsPlaceholder": "— Ścieżek",
"footer.stemsCount.one": "{count} ścieżka",
@@ -1457,7 +1476,7 @@ const ja = {
"sections.clearConfirm": "確認?",
"structure.toggle": "曲の構成",
"structure.toggleTitle": "実験的な楽曲構造抽出: 分離後にイントロ、Aメロ、サビを自動でラベル付けします",
- "structure.experimental": "試験的",
+ "common.experimental": "試験的",
"doc.title": "StemDeck — トラックをパートごとに分離",
"topbar.urlPlaceholder": "検索するか、音声ファイルをドロップ…",
@@ -1471,6 +1490,8 @@ const ja = {
"stem.vocals": "ボーカル",
"stem.lead_vocals": "リードボーカル",
"stem.backing_vocals": "コーラス",
+ "stem.voice_1": "ボイス 1",
+ "stem.voice_2": "ボイス 2",
"stem.drums": "ドラム",
"stem.bass": "ベース",
"stem.guitar": "ギター",
@@ -1481,6 +1502,8 @@ const ja = {
"vocalMode.groupAria": "ボーカルモード",
"vocalMode.all": "すべて",
"vocalMode.split": "リード+コーラス",
+ "vocalMode.duet": "デュエット",
+ "vocalMode.duetTitle": "実験的な機能です。ボーカルを2人の歌手に分割します。交互に歌い、声質が異なるときに最も有効です。同時に歌う部分では、もう一方の声が残ります。",
"aria.mute": "{name}をミュート",
"aria.solo": "{name}をソロ",
@@ -1603,6 +1626,8 @@ const ja = {
"wave.loadingAria": "波形を読み込み中",
"job.cancelAria": "ジョブをキャンセル",
"job.cancel": "キャンセル",
+ "job.splittingVocals": "リードとコーラスを分離中…",
+ "job.splittingDuet": "デュエットの声を分離中…",
"footer.stemsPlaceholder": "— パート",
"footer.stemsCount.other": "{count}パート",
@@ -2039,7 +2064,7 @@ const zhHans = {
"sections.clearConfirm": "确认?",
"structure.toggle": "歌曲结构",
"structure.toggleTitle": "实验性歌曲结构提取: 分离后自动标记前奏、主歌和副歌",
- "structure.experimental": "实验性",
+ "common.experimental": "实验性",
"doc.title": "StemDeck — 将任意曲目分离为音轨",
"topbar.urlPlaceholder": "搜索，或拖放音频文件…",
@@ -2053,6 +2078,8 @@ const zhHans = {
"stem.vocals": "人声",
"stem.lead_vocals": "主唱",
"stem.backing_vocals": "和声",
+ "stem.voice_1": "声部 1",
+ "stem.voice_2": "声部 2",
"stem.drums": "鼓组",
"stem.bass": "贝斯",
"stem.guitar": "吉他",
@@ -2063,6 +2090,8 @@ const zhHans = {
"vocalMode.groupAria": "人声模式",
"vocalMode.all": "全部",
"vocalMode.split": "主唱+和声",
+ "vocalMode.duet": "二重唱",
+ "vocalMode.duetTitle": "实验性功能。将人声分成两位歌手。当他们轮流演唱且音色不同时效果最佳。两人同时演唱时，仍会听到另一个声音。",
"aria.mute": "静音{name}",
"aria.solo": "独奏{name}",
@@ -2185,6 +2214,8 @@ const zhHans = {
"wave.loadingAria": "正在加载波形",
"job.cancelAria": "取消任务",
"job.cancel": "取消",
+ "job.splittingVocals": "正在分离主唱与和声…",
+ "job.splittingDuet": "正在分离二重唱声部…",
"footer.stemsPlaceholder": "— 音轨",
"footer.stemsCount.other": "{count} 条音轨",
@@ -2621,7 +2652,7 @@ const de = {
"sections.clearConfirm": "Sicher?",
"structure.toggle": "Songstruktur",
"structure.toggleTitle": "Experimentelle Songstruktur-Erkennung: beschriftet Intro, Strophe und Refrain nach dem Trennen automatisch",
- "structure.experimental": "Experimentell",
+ "common.experimental": "EXPERIMENTELL",
"doc.title": "StemDeck — jeden Track in Stems zerlegen",
"topbar.urlPlaceholder": "Suchen oder eine Audiodatei ablegen…",
@@ -2635,6 +2666,8 @@ const de = {
"stem.vocals": "Gesang",
"stem.lead_vocals": "Lead-Gesang",
"stem.backing_vocals": "Backing-Gesang",
+ "stem.voice_1": "Stimme 1",
+ "stem.voice_2": "Stimme 2",
"stem.drums": "Schlagzeug",
"stem.bass": "Bass",
"stem.guitar": "Gitarre",
@@ -2645,6 +2678,8 @@ const de = {
"vocalMode.groupAria": "Gesangsmodus",
"vocalMode.all": "Alle",
"vocalMode.split": "Lead + Backing",
+ "vocalMode.duet": "Duette",
+ "vocalMode.duetTitle": "EXPERIMENTELL. Teilt den Gesang in zwei Stimmen auf. Funktioniert am besten, wenn sie sich abwechseln und unterschiedlich klingen. Wo beide gleichzeitig singen, bleibt etwas der anderen Stimme.",
"aria.mute": "{name} stummschalten",
"aria.solo": "{name} solo",
@@ -2767,6 +2802,8 @@ const de = {
"wave.loadingAria": "Wellenform wird geladen",
"job.cancelAria": "Auftrag abbrechen",
"job.cancel": "Abbrechen",
+ "job.splittingVocals": "Lead/Backing werden getrennt…",
+ "job.splittingDuet": "Duett-Stimmen werden getrennt…",
"footer.stemsPlaceholder": "— Stems",
"footer.stemsCount": "{count} Stems",
@@ -3214,7 +3251,7 @@ const pt = {
"sections.clearConfirm": "Confirmar?",
"structure.toggle": "Estrutura da música",
"structure.toggleTitle": "Extração experimental da estrutura da música: marca intro, verso e refrão automaticamente após a separação",
- "structure.experimental": "Experimental",
+ "common.experimental": "EXPERIMENTAL",
"doc.title": "StemDeck — separe qualquer faixa em stems",
"topbar.urlPlaceholder": "Pesquise ou solte um arquivo de áudio…",
@@ -3228,6 +3265,8 @@ const pt = {
"stem.vocals": "Vocais",
"stem.lead_vocals": "Vocal principal",
"stem.backing_vocals": "Vocal de apoio",
+ "stem.voice_1": "Voz 1",
+ "stem.voice_2": "Voz 2",
"stem.drums": "Bateria",
"stem.bass": "Baixo",
"stem.guitar": "Guitarra",
@@ -3238,6 +3277,8 @@ const pt = {
"vocalMode.groupAria": "Modo de vocais",
"vocalMode.all": "Todos",
"vocalMode.split": "Principal + Apoio",
+ "vocalMode.duet": "Duetos",
+ "vocalMode.duetTitle": "EXPERIMENTAL. Divide as vozes em dois cantores. Funciona melhor quando alternam e têm vozes diferentes. Onde cantam ao mesmo tempo, fica alguma da outra voz.",
"aria.mute": "Mudo {name}",
"aria.solo": "Solo {name}",
@@ -3360,6 +3401,8 @@ const pt = {
"wave.loadingAria": "Carregando forma de onda",
"job.cancelAria": "Cancelar tarefa",
"job.cancel": "Cancelar",
+ "job.splittingVocals": "A separar vocal principal e apoio…",
+ "job.splittingDuet": "A separar as vozes do dueto…",
"footer.stemsPlaceholder": "— Stems",
"footer.stemsCount": "{count} stems",
@@ -3809,7 +3852,7 @@ const id = {
"sections.clearConfirm": "Yakin?",
"structure.toggle": "Struktur lagu",
"structure.toggleTitle": "Ekstraksi struktur lagu eksperimental: menandai intro, bait, dan refrein secara otomatis setelah pemisahan",
- "structure.experimental": "Eksperimental",
+ "common.experimental": "EKSPERIMENTAL",
"doc.title": "StemDeck — pisahkan trek apa pun menjadi stem",
"topbar.urlPlaceholder": "Cari atau seret file audio…",
@@ -3823,6 +3866,8 @@ const id = {
"stem.vocals": "Vokal",
"stem.lead_vocals": "Vokal Utama",
"stem.backing_vocals": "Vokal Latar",
+ "stem.voice_1": "Suara 1",
+ "stem.voice_2": "Suara 2",
"stem.drums": "Drum",
"stem.bass": "Bass",
"stem.guitar": "Gitar",
@@ -3833,6 +3878,8 @@ const id = {
"vocalMode.groupAria": "Mode vokal",
"vocalMode.all": "Semua",
"vocalMode.split": "Utama + Latar",
+ "vocalMode.duet": "Duet",
+ "vocalMode.duetTitle": "EKSPERIMENTAL. Memisahkan vokal menjadi dua penyanyi. Paling baik bila mereka bergantian dan suaranya berbeda. Bila keduanya bernyanyi bersamaan, sebagian suara lain tetap terdengar.",
"aria.mute": "Bisukan {name}",
"aria.solo": "Solo {name}",
@@ -3955,6 +4002,8 @@ const id = {
"wave.loadingAria": "Memuat gelombang",
"job.cancelAria": "Batalkan tugas",
"job.cancel": "Batal",
+ "job.splittingVocals": "Memisahkan vokal utama/latar…",
+ "job.splittingDuet": "Memisahkan suara duet…",
"footer.stemsPlaceholder": "— Stem",
"footer.stemsCount": "{count} stem",
@@ -4391,7 +4440,7 @@ const fr = {
"sections.clearConfirm": "Confirmer ?",
"structure.toggle": "Structure du morceau",
"structure.toggleTitle": "Extraction expérimentale de la structure: étiquette automatiquement intro, couplet et refrain après la séparation",
- "structure.experimental": "Expérimental",
+ "common.experimental": "EXPÉRIMENTAL",
"doc.title": "StemDeck — séparez n'importe quel morceau en pistes",
"topbar.urlPlaceholder": "Recherchez ou déposez un fichier audio…",
@@ -4405,6 +4454,8 @@ const fr = {
"stem.vocals": "Voix",
"stem.lead_vocals": "Voix principale",
"stem.backing_vocals": "Chœurs",
+ "stem.voice_1": "Voix 1",
+ "stem.voice_2": "Voix 2",
"stem.drums": "Batterie",
"stem.bass": "Basse",
"stem.guitar": "Guitare",
@@ -4415,6 +4466,8 @@ const fr = {
"vocalMode.groupAria": "Mode voix",
"vocalMode.all": "Tout",
"vocalMode.split": "Principale + chœurs",
+ "vocalMode.duet": "Duos",
+ "vocalMode.duetTitle": "EXPÉRIMENTAL. Sépare le chant en deux interprètes. Fonctionne mieux quand ils alternent et que leurs voix diffèrent. Là où ils chantent ensemble, une partie de l'autre voix subsiste.",
"aria.mute": "Couper {name}",
"aria.solo": "Solo {name}",
@@ -4537,6 +4590,8 @@ const fr = {
"wave.loadingAria": "Chargement de la forme d'onde",
"job.cancelAria": "Annuler la tâche",
"job.cancel": "Annuler",
+ "job.splittingVocals": "Séparation voix principale/chœurs…",
+ "job.splittingDuet": "Séparation des voix du duo…",
"footer.stemsPlaceholder": "— Pistes",
"footer.stemsCount": "{count} pistes",
@@ -5102,7 +5157,7 @@ const es = {
"sections.clearConfirm": "¿Confirmar?",
"structure.toggle": "Estructura de la canción",
"structure.toggleTitle": "Extracción experimental de la estructura: etiqueta automáticamente intro, verso y estribillo tras la separación",
- "structure.experimental": "Experimental",
+ "common.experimental": "EXPERIMENTAL",
"doc.title": "StemDeck — separa cualquier pista en stems",
"topbar.urlPlaceholder": "Busca o suelta un archivo de audio…",
@@ -5116,6 +5171,8 @@ const es = {
"stem.vocals": "Voces",
"stem.lead_vocals": "Voz principal",
"stem.backing_vocals": "Coros",
+ "stem.voice_1": "Voz 1",
+ "stem.voice_2": "Voz 2",
"stem.drums": "Batería",
"stem.bass": "Bajo",
"stem.guitar": "Guitarra",
@@ -5126,6 +5183,8 @@ const es = {
"vocalMode.groupAria": "Modo de voces",
"vocalMode.all": "Todo",
"vocalMode.split": "Principal + coros",
+ "vocalMode.duet": "Dúos",
+ "vocalMode.duetTitle": "EXPERIMENTAL. Divide la voz en dos cantantes. Funciona mejor cuando se alternan y sus voces son distintas. Donde cantan a la vez, queda algo de la otra voz.",
"aria.mute": "Mute {name}",
"aria.solo": "Solo {name}",
@@ -5248,6 +5307,8 @@ const es = {
"wave.loadingAria": "Cargando forma de onda",
"job.cancelAria": "Cancelar tarea",
"job.cancel": "Cancelar",
+ "job.splittingVocals": "Separando voz principal y coros…",
+ "job.splittingDuet": "Separando las voces del dúo…",
"footer.stemsPlaceholder": "— Stems",
"footer.stemsCount": "{count} stems",
@@ -5718,7 +5779,7 @@ const ko = {
"sections.clearConfirm": "정말요?",
"structure.toggle": "곡 구조",
"structure.toggleTitle": "실험적 곡 구조 분석: 분리한 뒤 인트로, 벌스, 코러스를 자동으로 표시해요",
- "structure.experimental": "실험적",
+ "common.experimental": "실험적",
"doc.title": "StemDeck, 어떤 곡이든 스템으로 분리",
"topbar.urlPlaceholder": "검색하거나 오디오 파일을 끌어다 놓으세요…",
@@ -5732,6 +5793,8 @@ const ko = {
"stem.vocals": "보컬",
"stem.lead_vocals": "리드 보컬",
"stem.backing_vocals": "백 보컬",
+ "stem.voice_1": "목소리 1",
+ "stem.voice_2": "목소리 2",
"stem.drums": "드럼",
"stem.bass": "베이스",
"stem.guitar": "기타",
@@ -5742,6 +5805,8 @@ const ko = {
"vocalMode.groupAria": "보컬 모드",
"vocalMode.all": "전체",
"vocalMode.split": "리드 + 백",
+ "vocalMode.duet": "듀엣",
+ "vocalMode.duetTitle": "실험적 기능입니다. 보컬을 두 명의 가수로 나눕니다. 번갈아 부르고 목소리가 다를 때 가장 잘 작동합니다. 함께 부르는 부분에서는 다른 목소리가 일부 남습니다.",
"aria.mute": "{name} 음소거",
"aria.solo": "{name} 솔로",
@@ -5864,6 +5929,8 @@ const ko = {
"wave.loadingAria": "파형 불러오는 중",
"job.cancelAria": "작업 취소",
"job.cancel": "취소",
+ "job.splittingVocals": "리드/백 보컬 분리 중…",
+ "job.splittingDuet": "듀엣 목소리 분리 중…",
"footer.stemsPlaceholder": "— 스템",
"footer.stemsCount.other": "스템 {count}개",
diff --git a/static/js/job.js b/static/js/job.js
index 8ec73301..f337faf9 100644
--- a/static/js/job.js
+++ b/static/js/job.js
@@ -45,14 +45,21 @@ const splitAutoTriggered = new Set();
// than assuming their own `state` is still current.
export async function runVocalSplitIfWanted(state) {
if (state.status !== "done" || splitAutoTriggered.has(state.job_id)) return state;
- const wantsSplit = jobVocalSplitModes.get(state.job_id) === "split";
+ const mode = jobVocalSplitModes.get(state.job_id);
+ // The two decompositions are alternatives, so one table drives both: which
+ // endpoint to call, which job field says it already ran, and what to show
+ // while it runs.
+ const SPLITS = {
+ split: { path: "vocal-split", done: state.vocal_split, busy: t("job.splittingVocals") },
+ duet: { path: "duet-split", done: state.duet_split, busy: t("job.splittingDuet") },
+ };
+ const split = SPLITS[mode];
const hasVocals = (state.stems || []).some((s) => s.name === "vocals");
- const alreadySplit = state.vocal_split === "done";
- if (!wantsSplit || !hasVocals || alreadySplit) return state;
+ if (!split || !hasVocals || split.done === "done") return state;
splitAutoTriggered.add(state.job_id);
- if (state.job_id === foregroundJobId) setWaveformLoading(true, "Splitting lead/backing vocals…");
+ if (state.job_id === foregroundJobId) setWaveformLoading(true, split.busy);
try {
- const res = await fetch(`/api/jobs/${state.job_id}/vocal-split`, { method: "POST" });
+ const res = await fetch(`/api/jobs/${state.job_id}/${split.path}`, { method: "POST" });
if (!(res.ok || res.status === 202)) return state;
const r2 = await fetch(`/api/jobs/${state.job_id}`);
if (!r2.ok) return state;
diff --git a/static/js/mixer.js b/static/js/mixer.js
index 10b84314..b159186a 100644
--- a/static/js/mixer.js
+++ b/static/js/mixer.js
@@ -1,16 +1,18 @@
import {
- STEM_NAMES, TRACK_NAMES, EXTRA_STEM_NAMES, STEM_COLORS, STEM_DISPLAY, LANE_VOLUME_MAX,
+ STEM_NAMES, TRACK_NAMES, EXTRA_STEM_NAMES, DUET_STEM_NAMES, STEM_COLORS, STEM_DISPLAY,
+ LANE_VOLUME_MAX,
} from "./constants.js";
// Every lane name this session might ever need mixer state for, including
-// the on-demand lead/backing vocal split (#275). Safe to iterate broadly:
+// both on-demand vocal splits (#275 lead/backing, and the duet pair). Safe
+// to iterate broadly:
// callers below all guard on the row/trackIndex entry actually existing, so
// touching state for a lane a given job doesn't have is a no-op. A function
// (not a snapshot) because STEM_NAMES/EXTRA_STEM_NAMES are reassigned once
// syncStemNamesFromAPI() resolves -- a const array here would freeze the
// fallback values from before that happens.
function allTrackNames() {
- return [...TRACK_NAMES, ...EXTRA_STEM_NAMES];
+ return [...TRACK_NAMES, ...EXTRA_STEM_NAMES, ...DUET_STEM_NAMES];
}
import {
mixerState, mixerEl, stemListEl, currentJobId, multitrack, trackIndex,
diff --git a/static/js/playbackStems.js b/static/js/playbackStems.js
index a0342dee..29343481 100644
--- a/static/js/playbackStems.js
+++ b/static/js/playbackStems.js
@@ -5,6 +5,8 @@
// any drums inside it. Current jobs retain the individual Demucs stems, so the
// engine can rebuild that lane as a control group and route drums separately.
+import { EXTRA_STEM_NAMES, DUET_STEM_NAMES } from "./constants.js";
+
function playable(stem) {
return !!stem?.name && !!stem?.url;
}
@@ -36,10 +38,11 @@ export function buildPlaybackStems(rawStems, visibleStems, baseStemNames) {
if (!visible.some((stem) => stem.name === "original")) return result;
const dedicatedNames = new Set(dedicated.map((stem) => stem.name));
- // A completed lead/backing split replaces the base vocals source. Treat the
- // base stem as selected so it is not added to the complement as well.
- if (dedicatedNames.has("lead_vocals") && dedicatedNames.has("backing_vocals")) {
- dedicatedNames.add("vocals");
+ // A completed split (lead/backing, or the duet pair) replaces the base
+ // vocals source. Treat the base stem as selected so it is not added to the
+ // complement as well and counted twice.
+ for (const pair of [EXTRA_STEM_NAMES, DUET_STEM_NAMES]) {
+ if (pair.every((name) => dedicatedNames.has(name))) dedicatedNames.add("vocals");
}
const complementNames = baseStemNames.filter((name) => !dedicatedNames.has(name));
diff --git a/static/js/player.js b/static/js/player.js
index be2e32bc..120336e2 100644
--- a/static/js/player.js
+++ b/static/js/player.js
@@ -6,7 +6,7 @@ import { fmtTime } from "./utils.js";
// job.js state.
import { showPlaybackError, clearPlaybackError, resolvePlaybackSuccess } from "./job.js";
import {
- STEM_NAMES, TRACK_NAMES, EXTRA_STEM_NAMES, STEM_COLORS, PROGRESS_COLOR,
+ STEM_NAMES, TRACK_NAMES, EXTRA_STEM_NAMES, DUET_STEM_NAMES, STEM_COLORS, PROGRESS_COLOR,
LOOP_DEFAULT_START_FRAC, LOOP_DEFAULT_END_FRAC, LANE_VOLUME_MAX,
effectiveStemOrder,
} from "./constants.js";
@@ -95,14 +95,18 @@ const _STEM_ROW_SELECTORS = [
];
function applyStemSelectionFilter(presentNames) {
- // The order for THIS job: STEM_NAMES, with "vocals" swapped for
- // lead_vocals + backing_vocals when the on-demand split (#275) produced
- // both. Rows for a name outside `order` (the vocals-family row not in
- // play for this job) are hidden outright, not just grayed "unavailable" --
- // unlike the base 6, lead_vocals/backing_vocals aren't part of every job's
- // contract, so there's no "not selected this time" state for them to be in.
+ // The order for THIS job: STEM_NAMES, with "vocals" swapped for whichever
+ // on-demand split produced a complete pair -- lead_vocals + backing_vocals
+ // (#275), or the duet pair. Rows for a name outside `order` (the
+ // vocals-family rows not in play for this job) are hidden outright, not
+ // just grayed "unavailable" -- unlike the base 6, the split stems aren't
+ // part of every job's contract, so there's no "not selected this time"
+ // state for them to be in. Derived from the constants rather than listed
+ // here, so a new split pair does not silently render as an always-visible
+ // "unavailable" lane on every job.
const order = effectiveStemOrder(presentNames);
- const isVocalFamily = (s) => s === "vocals" || s === "lead_vocals" || s === "backing_vocals";
+ const vocalFamily = new Set(["vocals", ...EXTRA_STEM_NAMES, ...DUET_STEM_NAMES]);
+ const isVocalFamily = (s) => vocalFamily.has(s);
// Waveform rows: original hides if absent; order rows always show, grayed if absent
for (const el of document.querySelectorAll(".stem-waveform-row[data-stem]")) {
@@ -904,10 +908,11 @@ export function renderEmptyShell() {
stopStemVuLoop();
ensureMixerStateDefaults();
mixerEl.innerHTML = "";
- // lead_vocals/backing_vocals (#275) get rows too, built once here like the
- // base 6 -- applyStemSelectionFilter (below) hides them by default since no
- // job is loaded yet, and shows them in place of "vocals" once one is.
- for (const name of ["original", ...STEM_NAMES, ...EXTRA_STEM_NAMES]) {
+ // The split lanes (#275 lead/backing, and the duet pair) get rows too,
+ // built once here like the base 6 -- applyStemSelectionFilter (below) hides
+ // them by default since no job is loaded yet, and shows them in place of
+ // "vocals" once one is.
+ for (const name of ["original", ...STEM_NAMES, ...EXTRA_STEM_NAMES, ...DUET_STEM_NAMES]) {
const { row } = renderMixerRow({ name, url: "#" });
mixerEl.appendChild(row);
}
@@ -1118,17 +1123,24 @@ export function wireUpAudio(jobId, stems, duration, thumbnail, mixUrl = null, ti
// original.wav, so it's simply not in `stems` and the mixer/sidebar
// rows for it stay hidden.)
//
- // vocals/lead_vocals/backing_vocals are mutually exclusive (#275): once a
- // job's on-demand split has produced both, show those two in place of the
- // plain Vocals lane rather than all three -- vocals.wav is never deleted by
- // the split, so it would otherwise still show up here too.
+ // The vocals family is mutually exclusive (#275): once an on-demand split
+ // has produced a complete pair, show that pair in place of the plain Vocals
+ // lane rather than all of them -- vocals.wav is never deleted by a split, so
+ // it would otherwise still show up here too. The pair is whichever split
+ // ran, so it has to be looked up rather than named: a duet job's stems are
+ // not in selectedStems (which holds "vocals") and would be dropped by the
+ // final clause otherwise.
const rawPresent = new Set(stems.map((s) => s.name));
- const splitDone = rawPresent.has("lead_vocals") && rawPresent.has("backing_vocals");
+ const splitPair = [EXTRA_STEM_NAMES, DUET_STEM_NAMES].find((names) =>
+ names.every((n) => rawPresent.has(n))
+ );
const wantsVocals = selectedStems.has("vocals");
stems = stems.filter((s) => {
if (s.name === "original") return true;
- if (s.name === "vocals") return wantsVocals && !splitDone;
- if (s.name === "lead_vocals" || s.name === "backing_vocals") return wantsVocals && splitDone;
+ if (s.name === "vocals") return wantsVocals && !splitPair;
+ if (EXTRA_STEM_NAMES.includes(s.name) || DUET_STEM_NAMES.includes(s.name)) {
+ return wantsVocals && !!splitPair && splitPair.includes(s.name);
+ }
return selectedStems.has(s.name);
});
const playbackStems = buildPlaybackStems(rawStems, stems, STEM_NAMES);
diff --git a/static/js/state.js b/static/js/state.js
index e181d58d..e9a0a44c 100644
--- a/static/js/state.js
+++ b/static/js/state.js
@@ -158,20 +158,22 @@ export function setStemSelected(name, selected) {
saveSelectedStems();
}
-// On-demand lead/backing vocal split (#275): "all" (default, plain Vocals
-// lane) or "split" (auto-run the split once the next import finishes).
-// A page-level setting like selectedStems, applied at whatever moment the
-// user submits -- not stored per-job.
+// On-demand vocal decomposition (#275): "all" (default, plain Vocals lane),
+// "split" (lead + backing) or "duet" (two voices by register). The last two
+// are alternatives -- one extra model pass, not both. A page-level setting
+// like selectedStems, applied at whatever moment the user submits -- not
+// stored per-job.
const _VOCAL_SPLIT_MODE_KEY = "stemdeck:vocal-split-mode";
+const _VOCAL_SPLIT_MODES = ["split", "duet"];
export let vocalSplitMode = "all";
export const vocalSplitModeReady = (async () => {
try {
const v = await storeGet(_VOCAL_SPLIT_MODE_KEY, null);
- if (v === "split") vocalSplitMode = v;
+ if (_VOCAL_SPLIT_MODES.includes(v)) vocalSplitMode = v;
} catch (e) { console.warn("[state] failed to load vocal split mode:", e); }
})();
export function setVocalSplitMode(mode) {
- vocalSplitMode = mode === "split" ? "split" : "all";
+ vocalSplitMode = _VOCAL_SPLIT_MODES.includes(mode) ? mode : "all";
storeSet(_VOCAL_SPLIT_MODE_KEY, vocalSplitMode).catch((e) =>
console.warn("[state] failed to save vocal split mode:", e)
);
diff --git a/tests/e2e/file-drop.spec.mjs b/tests/e2e/file-drop.spec.mjs
index fed2f314..59d1a4e0 100644
--- a/tests/e2e/file-drop.spec.mjs
+++ b/tests/e2e/file-drop.spec.mjs
@@ -170,6 +170,37 @@ test.describe("file drop", () => {
await expect(page.locator("#url")).toBeVisible();
});
+ test("the refusal still has room to show on a small window", async ({ page }) => {
+ // The composer is one fixed-height row sharing a hard width budget, and the
+ // URL zone is the only item in it that flexes, so every pixel the stem
+ // controls ask for comes out of the zone. Adding a third vocal mode button
+ // took it to nothing but its own padding at 1280px: the input had no room,
+ // and neither did the message that sits in its place, so refusing a file
+ // answered into a zero-width box. Asserted at the narrowest width the
+ // desktop shell allows rather than at Playwright's default, because the
+ // default is what made this visible by luck.
+ await seedLibrary(page);
+ await stubTauri(page);
+ await stubUpdateCheck(page);
+ await page.setViewportSize({ width: 1024, height: 700 });
+ await page.goto("/", { waitUntil: "domcontentloaded" });
+ await page.waitForSelector(".url-wrap");
+
+ await dropOn(page, "body", "notes.txt", Buffer.from("hello"));
+ const err = page.locator("#urlDropError");
+ await expect(err).toBeVisible();
+ expect((await err.boundingBox()).width).toBeGreaterThan(40);
+
+ // And nothing was pushed out of the pill to pay for it: the button that
+ // starts the job is the one thing that must never be unreachable.
+ const fits = await page.evaluate(() => {
+ const c = document.querySelector(".daw-composer").getBoundingClientRect();
+ const b = document.querySelector(".daw-process-btn").getBoundingClientRect();
+ return b.right <= c.right + 1 && b.width > 0;
+ });
+ expect(fits, "the process button is inside the composer").toBe(true);
+ });
+
test("a library drag is not mistaken for a file drop", async ({ page }) => {
// The library's own drags carry no file list. If the document handler
// stopped guarding on that, dragging a track would arm the importer.
diff --git a/tests/test_cancellation_reach.py b/tests/test_cancellation_reach.py
index 8f0a79b8..09c303ec 100644
--- a/tests/test_cancellation_reach.py
+++ b/tests/test_cancellation_reach.py
@@ -103,7 +103,8 @@ def test_every_worker_spawn_exports_the_parent_pid():
for path in (
"app/pipeline/separate.py",
- "app/pipeline/vocal_split.py",
+ # The on-demand splits (lead/backing, duet) share one supervisor.
+ "app/pipeline/split_runner.py",
"app/pipeline/sections.py",
):
src = pathlib.Path(path).read_text()
@@ -116,6 +117,7 @@ def test_every_worker_arms_the_watchdog():
for path in (
"app/pipeline/demucs_worker.py",
"app/pipeline/vocal_split_worker.py",
+ "app/pipeline/duet_split_worker.py",
"app/pipeline/section_worker.py",
):
src = pathlib.Path(path).read_text()
diff --git a/tests/test_duet_medleyvox.py b/tests/test_duet_medleyvox.py
new file mode 100644
index 00000000..3643e34b
--- /dev/null
+++ b/tests/test_duet_medleyvox.py
@@ -0,0 +1,164 @@
+"""Tests for the same-register duet split.
+
+Two things are worth pinning down. The collapse check, because it is the one
+property of a split that can be judged without knowing the right answer and it
+decides whether a second, slower model runs. And the separator's output
+contract: stems that sum back to the recording, at the recording's own rate.
+
+The separator is exercised through a stub model, so the mask maths and the
+sample-rate transfer are covered without a 460 MB checkpoint in CI.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import pytest
+import soundfile as sf
+
+from app.pipeline.duet_medleyvox import MODEL_SR, register_split_collapsed, separate
+
+SR = 44100
+
+
+def _write(path, x, sr=SR):
+ sf.write(str(path), np.asarray(x, dtype=np.float32), sr)
+ return path
+
+
+def _tone(f, n, sr=SR, amp=1.0):
+ t = np.arange(n) / sr
+ return amp * np.sin(2 * np.pi * f * t)
+
+
+def test_collapse_into_one_stem_is_a_failure(tmp_path):
+ """The shape the register model takes on two singers of the same range."""
+ n = SR * 6
+ a = _write(tmp_path / "a.wav", _tone(220, n))
+ b = _write(tmp_path / "b.wav", _tone(220, n, amp=0.02))
+ assert register_split_collapsed(a, b)
+
+
+def test_alternating_split_passes(tmp_path):
+ """Two stems that trade the lead, as a separated duet does."""
+ n = SR * 8
+ gate = np.zeros(n)
+ for i in range(0, n, SR * 2):
+ gate[i : i + SR] = 1.0
+ a = _write(tmp_path / "a.wav", _tone(220, n) * gate)
+ b = _write(tmp_path / "b.wav", _tone(330, n) * (1.0 - gate))
+ assert not register_split_collapsed(a, b)
+
+
+def test_silence_counts_as_failure(tmp_path):
+ n = SR * 4
+ a = _write(tmp_path / "a.wav", np.zeros(n))
+ b = _write(tmp_path / "b.wav", np.zeros(n))
+ assert register_split_collapsed(a, b)
+
+
+class _StubModel:
+ """Stands in for the checkpoint: splits by frequency, nothing learned.
+
+ Lets the mask averaging and the 24 kHz-to-file-rate transfer be tested
+ without downloading 460 MB, which is the part with real arithmetic in it.
+ """
+
+ def to(self, device):
+ return self
+
+ def eval(self):
+ return self
+
+ def __call__(self, x):
+ import torch
+
+ wav = x[0, 0].numpy().astype(np.float64)
+ spec = np.fft.rfft(wav)
+ freqs = np.fft.rfftfreq(len(wav), 1 / MODEL_SR)
+ lo = np.fft.irfft(spec * (freqs < 400), n=len(wav))
+ hi = wav - lo
+ return torch.as_tensor(np.stack([lo, hi])[None, :, :], dtype=torch.float32)
+
+
+def test_separate_preserves_the_recording_and_its_rate():
+ """Whatever the model says, the two stems must add back to the input.
+
+ Mixture consistency plus the mask normalisation is what makes solo and mute
+ behave in the mixer, so it is worth asserting rather than assuming.
+ """
+ n = SR * 3
+ mix = (_tone(180, n) + _tone(900, n, amp=0.7)).reshape(-1, 1)
+ mix = mix / np.max(np.abs(mix))
+
+ stems = separate([_StubModel()], mix, SR, "cpu")
+
+ assert len(stems) == 2
+ for s in stems:
+ assert len(s) == n, "stems must come back at the input length"
+ assert np.all(np.isfinite(s))
+
+ mono = mix[:, 0]
+ residual = mono - (stems[0] + stems[1])
+ rel = np.sqrt(np.mean(residual**2)) / np.sqrt(np.mean(mono**2))
+ assert rel < 1e-6, f"stems do not sum to the input (relative residual {rel:.2e})"
+
+
+def test_separate_keeps_content_above_the_models_ceiling():
+ """The model is 24 kHz; the output must not be band-limited to 12 kHz.
+
+ Separating at 24 kHz and resampling back would silently drop the air and
+ most of the sibilance from a vocal. The mask is transferred to the file's
+ own grid instead, so energy above the model's ceiling has to survive.
+ """
+ n = SR * 3
+ mix = (_tone(300, n) + _tone(16000, n, amp=0.3)).reshape(-1, 1)
+ mix = mix / np.max(np.abs(mix))
+
+ stems = separate([_StubModel()], mix, SR, "cpu")
+ recombined = stems[0] + stems[1]
+
+ spec = np.abs(np.fft.rfft(recombined))
+ freqs = np.fft.rfftfreq(len(recombined), 1 / SR)
+ above = float(np.sum(spec[freqs > 12000] ** 2))
+ total = float(np.sum(spec**2)) + 1e-12
+ assert above / total > 0.01, "content above 12 kHz was lost"
+
+
+@pytest.mark.parametrize("channels", [1, 2])
+def test_separate_keeps_the_input_channel_count(channels):
+ """Stereo in, stereo out.
+
+ The model is mono, so the obvious implementation hands back mono stems while
+ every other stem in the job is stereo. That is not a cosmetic difference:
+ a mono stem among stereo ones loses its pan position and width in the mixer.
+ The mask is decided on the downmix and applied to each channel.
+ """
+ n = SR * 2
+ left = _tone(220, n) + _tone(660, n, amp=0.5)
+ right = _tone(220, n, amp=0.4) + _tone(660, n)
+ wav = left.reshape(-1, 1) if channels == 1 else np.stack([left, right], axis=1)
+
+ stems = separate([_StubModel()], wav, SR, "cpu")
+
+ for stem in stems:
+ assert len(stem) == n
+ if channels == 1:
+ assert stem.ndim == 1
+ else:
+ assert stem.shape == (n, 2)
+
+ # Per channel, the two stems must still add back to that channel.
+ for ch in range(channels):
+ src = wav[:, ch]
+ got = stems[0] + stems[1]
+ got = got if channels == 1 else got[:, ch]
+ rel = np.sqrt(np.mean((src - got) ** 2)) / np.sqrt(np.mean(src**2))
+ assert rel < 1e-6, f"channel {ch} does not reconstruct (residual {rel:.2e})"
+
+ if channels == 2:
+ # A real stereo image must survive: the two channels of a stem must not
+ # collapse to the same signal.
+ for stem in stems:
+ if np.sqrt(np.mean(stem[:, 0] ** 2)) > 1e-6:
+ diff = np.mean(np.abs(stem[:, 0] - stem[:, 1]))
+ assert diff > 1e-6, "stereo stem collapsed to dual mono"
diff --git a/tests/test_jobs_api.py b/tests/test_jobs_api.py
index 3a48d042..078e182b 100644
--- a/tests/test_jobs_api.py
+++ b/tests/test_jobs_api.py
@@ -821,3 +821,116 @@ def test_submit_captures_the_song_structure_setting_on_the_job(client, monkeypat
r = client.post("/api/jobs", json={"url": "https://youtu.be/oHg5SJYRHA0"})
assert r.status_code == 200
assert registry.get(r.json()["job_id"]).auto_sections is False
+
+
+# ─── POST /jobs/{id}/duet-split ──────────────────────────────────────────────
+#
+# Mirrors the vocal-split cases above. The two splits are alternatives, so the
+# last test here is the one that matters most: running one must not make the
+# other unavailable.
+
+
+def test_duet_split_404_unknown_job(client):
+ r = client.post("/api/jobs/000000000000/duet-split")
+ assert r.status_code == 404
+
+
+def test_duet_split_404_job_not_done(client, tmp_path, monkeypatch):
+ import app.api.jobs as jobs_mod
+
+ monkeypatch.setattr(jobs_mod, "JOBS_DIR", tmp_path)
+ job = Job(id="abcdefabc630", status="separating", title="In progress")
+ _jobs[job.id] = job
+
+ r = client.post(f"/api/jobs/{job.id}/duet-split")
+ assert r.status_code == 404
+
+
+def _fake_duet(job_arg, stems_dir_arg):
+ (stems_dir_arg / "voice_1.wav").write_bytes(b"RIFF")
+ (stems_dir_arg / "voice_2.wav").write_bytes(b"RIFF")
+ return ["voice_1", "voice_2"]
+
+
+def test_duet_split_success(client, tmp_path, monkeypatch):
+ import app.api.jobs as jobs_mod
+
+ job, _ = _make_done_job_with_vocals(tmp_path, monkeypatch, job_id="abcdefabc630")
+ monkeypatch.setattr(jobs_mod, "split_duet", _fake_duet)
+
+ r = client.post(f"/api/jobs/{job.id}/duet-split")
+ assert r.status_code == 200
+ body = r.json()
+ assert body["duet_split"] == "done"
+ names = {s["name"] for s in body["stems"]}
+ assert {"voice_1", "voice_2"} <= names
+ assert job.duet_split == "done"
+
+
+def test_duet_split_409_when_already_running(client, tmp_path, monkeypatch):
+ job, _ = _make_done_job_with_vocals(tmp_path, monkeypatch, job_id="abcdefabc630")
+ job.duet_split = "running"
+
+ r = client.post(f"/api/jobs/{job.id}/duet-split")
+ assert r.status_code == 409
+
+
+def test_duet_split_202_idempotent_when_already_done(client, tmp_path, monkeypatch):
+ import app.api.jobs as jobs_mod
+
+ job, _ = _make_done_job_with_vocals(tmp_path, monkeypatch, job_id="abcdefabc630")
+ job.duet_split = "done"
+ calls = []
+
+ def should_not_run(job_arg, stems_dir_arg):
+ calls.append(1)
+ return ["voice_1", "voice_2"]
+
+ monkeypatch.setattr(jobs_mod, "split_duet", should_not_run)
+
+ r = client.post(f"/api/jobs/{job.id}/duet-split")
+ assert r.status_code == 202
+ assert calls == [], "the expensive model must not re-run once the split is done"
+
+
+def test_duet_split_failure_keeps_job_done_with_base_stems(client, tmp_path, monkeypatch):
+ import app.api.jobs as jobs_mod
+ from app.pipeline.errors import SeparationError
+
+ job, stems_dir = _make_done_job_with_vocals(tmp_path, monkeypatch, job_id="abcdefabc630")
+
+ def boom(job_arg, stems_dir_arg):
+ raise SeparationError("duet split failed: model download failed")
+
+ monkeypatch.setattr(jobs_mod, "split_duet", boom)
+
+ r = client.post(f"/api/jobs/{job.id}/duet-split")
+ assert r.status_code == 500
+ assert job.status == "done", "a failed split must not fail the job"
+ assert job.duet_split == "error"
+ assert (stems_dir / "duet_split_error.txt").is_file()
+ assert (stems_dir / "vocals.wav").is_file(), "base stems must be untouched"
+
+
+def test_duet_split_and_vocal_split_are_independent(client, tmp_path, monkeypatch):
+ """Running one split must not block or overwrite the other."""
+ import app.api.jobs as jobs_mod
+
+ job, _ = _make_done_job_with_vocals(tmp_path, monkeypatch, job_id="abcdefabc630")
+
+ def fake_vocals(job_arg, stems_dir_arg):
+ (stems_dir_arg / "lead_vocals.wav").write_bytes(b"RIFF")
+ (stems_dir_arg / "backing_vocals.wav").write_bytes(b"RIFF")
+ return ["lead_vocals", "backing_vocals"]
+
+ monkeypatch.setattr(jobs_mod, "split_vocals", fake_vocals)
+ monkeypatch.setattr(jobs_mod, "split_duet", _fake_duet)
+
+ assert client.post(f"/api/jobs/{job.id}/vocal-split").status_code == 200
+ assert client.post(f"/api/jobs/{job.id}/duet-split").status_code == 200
+
+ body = client.get(f"/api/jobs/{job.id}").json()
+ names = {s["name"] for s in body["stems"]}
+ assert {"lead_vocals", "backing_vocals", "voice_1", "voice_2"} <= names
+ assert body["vocal_split"] == "done"
+ assert body["duet_split"] == "done"
diff --git a/tests/test_pipeline_duet_split.py b/tests/test_pipeline_duet_split.py
new file mode 100644
index 00000000..caa4c781
--- /dev/null
+++ b/tests/test_pipeline_duet_split.py
@@ -0,0 +1,106 @@
+"""Tests for the on-demand duet split.
+
+Mirrors test_pipeline_vocal_split.py: the real subprocess machinery (Popen,
+stderr streaming, watchdog, cleanup) runs end-to-end against a stub worker
+script swapped in via the _spawn_cmd seam, so no real ML model or GPU is
+needed.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+from app.core.models import Job
+from app.pipeline import duet_split as ds_mod
+from app.pipeline.errors import SeparationError
+
+_SUCCESS_WORKER = """
+import sys, os
+out_dir = sys.argv[3]
+open(os.path.join(out_dir, "voice_1.wav"), "wb").write(b"RIFF")
+open(os.path.join(out_dir, "voice_2.wav"), "wb").write(b"RIFF")
+sys.stderr.write("@@DONE@@\\n")
+sys.stderr.flush()
+"""
+
+_FAILING_WORKER = """
+import sys, json
+sys.stderr.write("@@ERROR@@" + json.dumps("model download failed: connection reset") + "\\n")
+sys.stderr.flush()
+sys.exit(1)
+"""
+
+# Exits 0 / prints @@DONE@@ but never writes the two output files -- a
+# worker-side bug this should still catch rather than trust blindly.
+_LIAR_WORKER = """
+import sys
+sys.stderr.write("@@DONE@@\\n")
+sys.stderr.flush()
+"""
+
+
+def _stub(code: str):
+ def fake_spawn(device: str, vocals_path: Path, out_dir: Path) -> list[str]:
+ return [sys.executable, "-c", code, device, str(vocals_path), str(out_dir)]
+
+ return fake_spawn
+
+
+@pytest.fixture()
+def stems_dir(tmp_path: Path) -> Path:
+ d = tmp_path / "stems"
+ d.mkdir()
+ (d / "vocals.wav").write_bytes(b"RIFF")
+ return d
+
+
+@pytest.fixture()
+def job() -> Job:
+ return Job(id="abcdefabc630")
+
+
+def test_split_duet_success(job, stems_dir, monkeypatch):
+ monkeypatch.setattr(ds_mod, "get_demucs_device", lambda: "cpu")
+ monkeypatch.setattr(ds_mod, "_spawn_cmd", _stub(_SUCCESS_WORKER))
+
+ result = ds_mod.split_duet(job, stems_dir)
+
+ assert result == ["voice_1", "voice_2"]
+ assert (stems_dir / "voice_1.wav").is_file()
+ assert (stems_dir / "voice_2.wav").is_file()
+
+
+def test_split_duet_missing_vocals(job, tmp_path, monkeypatch):
+ empty = tmp_path / "stems"
+ empty.mkdir()
+ with pytest.raises(SeparationError, match="vocals.wav not found"):
+ ds_mod.split_duet(job, empty)
+
+
+def test_split_duet_worker_failure_surfaces_stderr(job, stems_dir, monkeypatch):
+ monkeypatch.setattr(ds_mod, "get_demucs_device", lambda: "cpu")
+ monkeypatch.setattr(ds_mod, "_spawn_cmd", _stub(_FAILING_WORKER))
+
+ with pytest.raises(SeparationError) as exc:
+ ds_mod.split_duet(job, stems_dir)
+ assert "connection reset" in str(exc.value)
+
+
+def test_split_duet_rejects_worker_that_wrote_nothing(job, stems_dir, monkeypatch):
+ monkeypatch.setattr(ds_mod, "get_demucs_device", lambda: "cpu")
+ monkeypatch.setattr(ds_mod, "_spawn_cmd", _stub(_LIAR_WORKER))
+
+ with pytest.raises(SeparationError, match="did not produce"):
+ ds_mod.split_duet(job, stems_dir)
+
+
+def test_split_duet_leaves_base_stems_untouched(job, stems_dir, monkeypatch):
+ monkeypatch.setattr(ds_mod, "get_demucs_device", lambda: "cpu")
+ monkeypatch.setattr(ds_mod, "_spawn_cmd", _stub(_SUCCESS_WORKER))
+
+ ds_mod.split_duet(job, stems_dir)
+
+ assert (stems_dir / "vocals.wav").read_bytes() == b"RIFF"
diff --git a/tests/test_stems_api.py b/tests/test_stems_api.py
index d98e9b78..0ddd700e 100644
--- a/tests/test_stems_api.py
+++ b/tests/test_stems_api.py
@@ -183,10 +183,50 @@ def test_peaks_returns_json_for_done_job(client, tmp_path):
r = client.get(f"/api/jobs/{job.id}/stems/peaks.json")
assert r.status_code == 200
assert r.headers["content-type"] == "application/json"
- assert "immutable" in r.headers.get("cache-control", "")
assert r.json() == payload
+def test_peaks_are_never_served_immutable(client, tmp_path):
+ """The browser must always ask whether this file changed.
+
+ It used to be sent `immutable` for a year, which was true when a finished
+ job's stems never changed again. The on-demand splits broke that: both add
+ stems to a done job and rewrite peaks.json. `immutable` is taken literally,
+ so the browser would not revalidate even on a hard reload, and the user got
+ the new audio under the old waveform.
+ """
+ job = Job(id="abcdefabcde1")
+ job.status = "done"
+ _jobs[job.id] = job
+ _make_peaks_file(tmp_path, job.id, {"vocals": [[-0.1, 0.2]]})
+
+ cc = client.get(f"/api/jobs/{job.id}/stems/peaks.json").headers.get("cache-control", "")
+ assert "immutable" not in cc, f"peaks.json must stay revalidatable, got {cc!r}"
+ assert "must-revalidate" in cc or "no-cache" in cc
+
+
+def test_peaks_reflect_a_split_that_added_stems(client, tmp_path):
+ """A split rewrites peaks.json, and the next request must show the new stems."""
+ job = Job(id="abcdefabcde2")
+ job.status = "done"
+ _jobs[job.id] = job
+ _make_peaks_file(tmp_path, job.id, {"vocals": [[-0.1, 0.2]]})
+
+ first = client.get(f"/api/jobs/{job.id}/stems/peaks.json").json()
+ assert "voice_1" not in first
+
+ # what the duet split does to this file
+ _make_peaks_file(
+ tmp_path,
+ job.id,
+ {"vocals": [[-0.1, 0.2]], "voice_1": [[-0.3, 0.3]], "voice_2": [[-0.4, 0.4]]},
+ )
+
+ second = client.get(f"/api/jobs/{job.id}/stems/peaks.json").json()
+ assert "voice_1" in second and "voice_2" in second
+ assert second["voice_1"] == [[-0.3, 0.3]]
+
+
def test_peaks_404_when_file_missing(client, tmp_path):
job = Job(id="abcdefabcdeb")
job.status = "done"
diff --git a/tests/test_subprocess_encoding.py b/tests/test_subprocess_encoding.py
index f0806e4d..d036413c 100644
--- a/tests/test_subprocess_encoding.py
+++ b/tests/test_subprocess_encoding.py
@@ -22,7 +22,8 @@
SITES = [
("app/pipeline/separate.py", "the Demucs worker"),
- ("app/pipeline/vocal_split.py", "the vocal-split worker"),
+ # Both on-demand splits (lead/backing, duet) share one supervisor.
+ ("app/pipeline/split_runner.py", "the on-demand split workers"),
("app/api/jobs.py", "ffprobe on upload"),
]
@@ -35,7 +36,7 @@ def test_decoding_is_pinned_not_left_to_the_locale(path: str, what: str):
assert 'errors="replace"' in src, f"{what} would still raise on an undecodable byte"
-@pytest.mark.parametrize("path", ["app/pipeline/separate.py", "app/pipeline/vocal_split.py"])
+@pytest.mark.parametrize("path", ["app/pipeline/separate.py", "app/pipeline/split_runner.py"])
def test_children_are_told_to_write_utf8(path: str):
"""The parent reading utf-8 is only half of it."""
src = Path(path).read_text(encoding="utf-8")
diff --git a/uv.lock b/uv.lock
index e600d1ba..391e7cb5 100644
--- a/uv.lock
+++ b/uv.lock
@@ -85,35 +85,52 @@ wheels = [
{ url = "https://files.pythonhosted.org/packages/da/42/e921fccf5015463e32a3cf6ee7f980a6ed0f395ceeaa45060b61d86486c2/anyio-4.13.0-py3-none-any.whl", hash = "sha256:08b310f9e24a9594186fd75b4f73f4a4152069e3853f1ed8bfbf58369f4ad708", size = 114353, upload-time = "2026-03-24T12:59:08.246Z" },
]
+[[package]]
+name = "asteroid-filterbanks"
+version = "0.4.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "torch", version = "2.2.2", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/90/fa/5c2be1f96dc179f83cdd3bb267edbd1f47d08f756785c016d5c2163901a7/asteroid-filterbanks-0.4.0.tar.gz", hash = "sha256:415f89d1dcf2b13b35f03f7a9370968ac4e6fa6800633c522dac992b283409b9", size = 24599, upload-time = "2021-04-09T20:03:07.456Z" }
+wheels = [
+ { url = "https://files.pythonhosted.org/packages/c5/7c/83ff6046176a675e6a1e8aeefed8892cd97fe7c46af93cc540d1b24b8323/asteroid_filterbanks-0.4.0-py3-none-any.whl", hash = "sha256:4932ac8b6acc6e08fb87cbe8ece84215b5a74eee284fe83acf3540a72a02eaf5", size = 29912, upload-time = "2021-04-09T20:03:05.817Z" },
+]
+
[[package]]
name = "audio-separator"
version = "0.44.5"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "audioop-lts", marker = "python_full_version >= '3.13'" },
- { name = "beartype" },
- { name = "diffq", marker = "sys_platform != 'win32'" },
+ { name = "audioop-lts", marker = "(python_full_version >= '3.13' and platform_machine != 'x86_64') or (python_full_version >= '3.13' and sys_platform != 'darwin')" },
+ { name = "beartype", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "diffq", marker = "(platform_machine != 'x86_64' and sys_platform == 'darwin') or (sys_platform != 'darwin' and sys_platform != 'win32')" },
{ name = "diffq-fixed", marker = "sys_platform == 'win32'" },
- { name = "einops" },
- { name = "julius" },
- { name = "librosa" },
- { name = "ml-collections" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
- { name = "onnx-weekly" },
- { name = "onnx2torch-py313" },
- { name = "pydub" },
- { name = "pyyaml" },
- { name = "requests" },
- { name = "resampy" },
- { name = "rotary-embedding-torch", version = "0.6.5", source = { registry = "https://pypi.org/simple" } },
- { name = "samplerate" },
- { name = "scipy", version = "1.15.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "scipy", version = "1.17.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
- { name = "six" },
- { name = "soundfile" },
- { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" } },
- { name = "tqdm" },
+ { name = "einops", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "julius", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "librosa", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "ml-collections", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "onnx-weekly", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "onnx2torch-py313", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "pydub", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "pyyaml", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "requests", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "resampy", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "rotary-embedding-torch", version = "0.6.5", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "samplerate", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "scipy", version = "1.15.3", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "scipy", version = "1.17.1", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "six", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "soundfile", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "tqdm", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/e6/a1/91884a9cbbace7d1eb480f6446e10fc2cdf0aa8d86da414f4593557b6340/audio_separator-0.44.5.tar.gz", hash = "sha256:58866bc61d0c692fff8a52cda67c284f7847a844048987135677a007bf0ae794", size = 344654, upload-time = "2026-07-20T22:02:44.933Z" }
wheels = [
@@ -391,8 +408,8 @@ resolution-markers = [
"(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
]
sdist = { url = "https://files.pythonhosted.org/packages/66/54/eb9bfc647b19f2009dd5c7f5ec51c4e6ca831725f1aea7a993034f483147/contourpy-1.3.2.tar.gz", hash = "sha256:b6945942715a034c671b7fc54f9588126b0b8bf23db2696e3ca8328f3ff0ab54", size = 13466130, upload-time = "2025-04-15T17:47:53.79Z" }
wheels = [
@@ -466,8 +483,8 @@ resolution-markers = [
"(python_full_version == '3.11.*' and platform_machine != 'x86_64') or (python_full_version == '3.11.*' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
]
sdist = { url = "https://files.pythonhosted.org/packages/58/01/1253e6698a07380cd31a736d248a3f2a50a7c88779a1813da27503cadc2a/contourpy-1.3.3.tar.gz", hash = "sha256:083e12155b210502d0bca491432bb04d56dc3432f95a979b429f2848c3dbe880", size = 13466174, upload-time = "2025-07-26T12:03:12.549Z" }
wheels = [
@@ -621,10 +638,10 @@ name = "diffq"
version = "0.2.4"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "cython" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
- { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" } },
+ { name = "cython", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/5a/fd/4c58807bf855c5929ffa6da55f26dd6b9ae462a4193f5e09cc49fbbfd451/diffq-0.2.4.tar.gz", hash = "sha256:049064861e974ebf00d0badab8b324c775037371419eda3150985b9d477b5bd2", size = 157139, upload-time = "2023-05-05T12:39:43.089Z" }
wheels = [
@@ -640,10 +657,10 @@ name = "diffq-fixed"
version = "0.2.4"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "cython" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
- { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" } },
+ { name = "cython", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/be/96/8ca5acf5ecfd4108aa6f345cef171f2fda0c081cf0e7430671712586f172/diffq_fixed-0.2.4.tar.gz", hash = "sha256:cbc906b76fa23d1cf3c0ae517fbab744d9624980a068fe7fbb00dede1d83208d", size = 189695, upload-time = "2024-11-07T21:33:26.204Z" }
wheels = [
@@ -681,7 +698,7 @@ name = "exceptiongroup"
version = "1.3.1"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "typing-extensions" },
+ { name = "typing-extensions", marker = "python_full_version < '3.11'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/50/79/66800aadf48771f6b62f7eb014e352e5d06856655206165d775e675a02c9/exceptiongroup-1.3.1.tar.gz", hash = "sha256:8b412432c6055b0b7d14c310000ae93352ed6754f70fa8f7c34141f91c4e3219", size = 30371, upload-time = "2025-11-21T23:01:54.787Z" }
wheels = [
@@ -1256,16 +1273,16 @@ resolution-markers = [
"(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "contourpy", version = "1.3.2", source = { registry = "https://pypi.org/simple" } },
- { name = "cycler" },
- { name = "fonttools" },
- { name = "kiwisolver" },
- { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
- { name = "packaging" },
- { name = "pillow" },
- { name = "pyparsing" },
- { name = "python-dateutil" },
+ { name = "contourpy", version = "1.3.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
+ { name = "cycler", marker = "python_full_version < '3.11'" },
+ { name = "fonttools", marker = "python_full_version < '3.11'" },
+ { name = "kiwisolver", marker = "python_full_version < '3.11'" },
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "packaging", marker = "python_full_version < '3.11'" },
+ { name = "pillow", marker = "python_full_version < '3.11'" },
+ { name = "pyparsing", marker = "python_full_version < '3.11'" },
+ { name = "python-dateutil", marker = "python_full_version < '3.11'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/63/1b/4be5be87d43d327a0cf4de1a56e86f7f84c89312452406cf122efe2839e6/matplotlib-3.10.9.tar.gz", hash = "sha256:fd66508e8c6877d98e586654b608a0456db8d7e8a546eb1e2600efd957302358", size = 34811233, upload-time = "2026-04-24T00:14:13.539Z" }
wheels = [
@@ -1323,16 +1340,16 @@ resolution-markers = [
"(python_full_version == '3.11.*' and platform_machine != 'x86_64') or (python_full_version == '3.11.*' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "contourpy", version = "1.3.3", source = { registry = "https://pypi.org/simple" } },
- { name = "cycler" },
- { name = "fonttools" },
- { name = "kiwisolver" },
- { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
- { name = "packaging" },
- { name = "pillow" },
- { name = "pyparsing" },
- { name = "python-dateutil" },
+ { name = "contourpy", version = "1.3.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
+ { name = "cycler", marker = "python_full_version >= '3.11'" },
+ { name = "fonttools", marker = "python_full_version >= '3.11'" },
+ { name = "kiwisolver", marker = "python_full_version >= '3.11'" },
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "packaging", marker = "python_full_version >= '3.11'" },
+ { name = "pillow", marker = "python_full_version >= '3.11'" },
+ { name = "pyparsing", marker = "python_full_version >= '3.11'" },
+ { name = "python-dateutil", marker = "python_full_version >= '3.11'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/49/64/f9a391af28f518b11ad45a8a712353c94a0aefce09d3703200e5c54b610a/matplotlib-3.11.1.tar.gz", hash = "sha256:69647db5746941c793d6e445a4cd349323ffb87d9cc958c2ad84a659b4832d30", size = 32612045, upload-time = "2026-07-18T03:39:46.63Z" }
wheels = [
@@ -1374,8 +1391,8 @@ name = "ml-collections"
version = "1.1.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "absl-py" },
- { name = "pyyaml" },
+ { name = "absl-py", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "pyyaml", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/b8/f8/1a9ae6696dbb6bc9c44ddf5c5e84710d77fe9a35a57e8a06722e1836a4a6/ml_collections-1.1.0.tar.gz", hash = "sha256:0ac1ac6511b9f1566863e0bb0afad0c64e906ea278ad3f4d2144a55322671f6f", size = 61356, upload-time = "2025-04-17T08:25:02.247Z" }
wheels = [
@@ -1387,8 +1404,8 @@ name = "ml-dtypes"
version = "0.6.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
]
sdist = { url = "https://files.pythonhosted.org/packages/12/72/307d7c4bd0600601c7133fba5cb78af7db968152951c1cd473abb1cda782/ml_dtypes-0.6.0.tar.gz", hash = "sha256:5e60251d32ced5598972e4d5e06a2f044341f9291402551a3f6f0ec44f9299b0", size = 3032327, upload-time = "2026-08-13T14:14:40.215Z" }
wheels = [
@@ -1512,8 +1529,8 @@ resolution-markers = [
"python_full_version < '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'",
]
dependencies = [
- { name = "llvmlite", version = "0.44.0", source = { registry = "https://pypi.org/simple" } },
- { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" } },
+ { name = "llvmlite", version = "0.44.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/1c/a0/e21f57604304aa03ebb8e098429222722ad99176a4f979d34af1d1ee80da/numba-0.61.2.tar.gz", hash = "sha256:8750ee147940a6637b80ecf7f95062185ad8726c8c28a2295b8ec1160a196f7d", size = 2820615, upload-time = "2025-04-09T02:58:07.659Z" }
wheels = [
@@ -1534,9 +1551,9 @@ resolution-markers = [
"(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "llvmlite", version = "0.47.0", source = { registry = "https://pypi.org/simple" } },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
+ { name = "llvmlite", version = "0.47.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
]
sdist = { url = "https://files.pythonhosted.org/packages/f6/c5/db2ac3685833d626c0dcae6bd2330cd68433e1fd248d15f70998160d3ad7/numba-0.65.1.tar.gz", hash = "sha256:19357146c32fe9ed25059ab915e8465fb13951cf6b0aace3826b76886373ab23", size = 2765600, upload-time = "2026-04-24T02:02:56.551Z" }
wheels = [
@@ -1739,7 +1756,7 @@ name = "nvidia-cudnn-cu12"
version = "9.1.0.70"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "nvidia-cublas-cu12" },
+ { name = "nvidia-cublas-cu12", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/9f/fd/713452cd72343f682b1c7b9321e23829f00b842ceaedcda96e742ea0b0b3/nvidia_cudnn_cu12-9.1.0.70-py3-none-manylinux2014_x86_64.whl", hash = "sha256:165764f44ef8c61fcdfdfdbe769d687e06374059fbb388b6c89ecb0e28793a6f", size = 664752741, upload-time = "2024-04-22T15:24:15.253Z" },
@@ -1750,7 +1767,7 @@ name = "nvidia-cufft-cu12"
version = "11.2.1.3"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "nvidia-nvjitlink-cu12" },
+ { name = "nvidia-nvjitlink-cu12", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/27/94/3266821f65b92b3138631e9c8e7fe1fb513804ac934485a8d05776e1dd43/nvidia_cufft_cu12-11.2.1.3-py3-none-manylinux2014_x86_64.whl", hash = "sha256:f083fc24912aa410be21fa16d157fed2055dab1cc4b6934a0e03cba69eb242b9", size = 211459117, upload-time = "2024-04-03T20:57:40.402Z" },
@@ -1769,9 +1786,9 @@ name = "nvidia-cusolver-cu12"
version = "11.6.1.9"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "nvidia-cublas-cu12" },
- { name = "nvidia-cusparse-cu12" },
- { name = "nvidia-nvjitlink-cu12" },
+ { name = "nvidia-cublas-cu12", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "nvidia-cusparse-cu12", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "nvidia-nvjitlink-cu12", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/3a/e1/5b9089a4b2a4790dfdea8b3a006052cfecff58139d5a4e34cb1a51df8d6f/nvidia_cusolver_cu12-11.6.1.9-py3-none-manylinux2014_x86_64.whl", hash = "sha256:19e33fa442bcfd085b3086c4ebf7e8debc07cfe01e11513cc6d332fd918ac260", size = 127936057, upload-time = "2024-04-03T20:58:28.735Z" },
@@ -1782,7 +1799,7 @@ name = "nvidia-cusparse-cu12"
version = "12.3.1.170"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "nvidia-nvjitlink-cu12" },
+ { name = "nvidia-nvjitlink-cu12", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/db/f7/97a9ea26ed4bbbfc2d470994b8b4f338ef663be97b8f677519ac195e113d/nvidia_cusparse_cu12-12.3.1.170-py3-none-manylinux2014_x86_64.whl", hash = "sha256:ea4f11a2904e2a8dc4b1833cc1b5181cde564edd0d5cd33e3c168eff2d1863f1", size = 207454763, upload-time = "2024-04-03T20:58:59.995Z" },
@@ -1838,11 +1855,11 @@ name = "onnx-weekly"
version = "1.23.0.dev20260817"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "ml-dtypes" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
- { name = "protobuf" },
- { name = "typing-extensions" },
+ { name = "ml-dtypes", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "protobuf", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "typing-extensions", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/9e/44/f24ce74b129bbcbeac50e07e30ee807f3abaa711f81873fe2afe110e484d/onnx_weekly-1.23.0.dev20260817.tar.gz", hash = "sha256:860d42175f474562fe05236f168ba2340c66b566119750792caa8cada18c74da", size = 5983184, upload-time = "2026-08-17T00:28:03.157Z" }
wheels = [
@@ -1871,11 +1888,11 @@ name = "onnx2torch-py313"
version = "1.6.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
- { name = "onnx-weekly" },
- { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" } },
- { name = "torchvision" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "onnx-weekly", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "torchvision", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/1e/bb/6301c1ea4c8f8859aff60d2d58a038fd68f9cdb507d1c60776d4e3b93b1f/onnx2torch_py313-1.6.0.tar.gz", hash = "sha256:d27b54c7e170f12ad252fe186e6cab8f3cade8ee4e596682acc5a3afddadb753", size = 49594, upload-time = "2025-04-11T03:52:32.88Z" }
wheels = [
@@ -1890,11 +1907,11 @@ resolution-markers = [
"(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "flatbuffers" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" } },
- { name = "packaging" },
- { name = "protobuf" },
- { name = "sympy" },
+ { name = "flatbuffers", marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "packaging", marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "protobuf", marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "sympy", marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/15/41/3253db975a90c3ce1d475e2a230773a21cd7998537f0657947df6fb79861/onnxruntime-1.24.3-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:3e6456801c66b095c5cd68e690ca25db970ea5202bd0c5b84a2c3ef7731c5a3c", size = 17332766, upload-time = "2026-03-05T17:18:59.714Z" },
@@ -1926,10 +1943,10 @@ resolution-markers = [
"(python_full_version == '3.11.*' and platform_machine != 'x86_64') or (python_full_version == '3.11.*' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "flatbuffers" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" } },
- { name = "packaging" },
- { name = "protobuf" },
+ { name = "flatbuffers", marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "packaging", marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "protobuf", marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/7c/a8/0520890321b8ff40b908cf165a93eb58fbc8f85c14db637277ea866c9544/onnxruntime-1.29.0-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:07c5907474dec4a2792fd7626b753dc66707808385a6d9eecf993db0066a9d0f", size = 21420890, upload-time = "2026-08-17T22:53:33.429Z" },
@@ -2359,9 +2376,9 @@ name = "resampy"
version = "0.4.3"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "numba", version = "0.65.1", source = { registry = "https://pypi.org/simple" } },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
+ { name = "numba", version = "0.65.1", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
]
sdist = { url = "https://files.pythonhosted.org/packages/29/f1/34be702a69a5d272e844c98cee82351f880985cfbca0cc86378011078497/resampy-0.4.3.tar.gz", hash = "sha256:a0d1c28398f0e55994b739650afef4e3974115edbe96cd4bb81968425e916e47", size = 3080604, upload-time = "2024-03-05T20:36:08.119Z" }
wheels = [
@@ -2388,8 +2405,8 @@ resolution-markers = [
"(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "einops" },
- { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" } },
+ { name = "einops", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/99/7e/8dbb4a950a99ec2f459ec1eb73818c519b56a1fee8a9a98f7a57460a0d5a/rotary_embedding_torch-0.6.5.tar.gz", hash = "sha256:a3623274c559e0215922edc7cd14068d5a64dea1ca469e1ecd642b293ee8af78", size = 7114, upload-time = "2024-08-20T20:52:44.155Z" }
wheels = [
@@ -2406,8 +2423,8 @@ resolution-markers = [
"python_full_version < '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'",
]
dependencies = [
- { name = "einops" },
- { name = "torch", version = "2.2.2", source = { registry = "https://pypi.org/simple" } },
+ { name = "einops", marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "torch", version = "2.2.2", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/71/38/74783585b1f0282fddd3faf1abd6dd20977255c27e31737eced2d7ec05f1/rotary_embedding_torch-0.8.9.tar.gz", hash = "sha256:b213f153cad1d108064d930544fb3af678d56515893d3f869a7a146f87997e3f", size = 7497, upload-time = "2025-07-27T01:26:14.675Z" }
wheels = [
@@ -2444,9 +2461,9 @@ name = "samplerate"
version = "0.1.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "cffi" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
+ { name = "cffi", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
]
sdist = { url = "https://files.pythonhosted.org/packages/da/9c/6a13fbc59b1ceaef194a53d3aeafa983057133fb13660c423854184eafa7/samplerate-0.1.0.tar.gz", hash = "sha256:75ef725e6cd9c4545569caf4c47147beab7b53b2c36e5122e8c285d348f88847", size = 4044998, upload-time = "2017-02-24T00:01:56.505Z" }
wheels = [
@@ -2462,11 +2479,11 @@ resolution-markers = [
"(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "joblib" },
- { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
- { name = "scipy", version = "1.15.3", source = { registry = "https://pypi.org/simple" } },
- { name = "threadpoolctl" },
+ { name = "joblib", marker = "python_full_version < '3.11'" },
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "scipy", version = "1.15.3", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
+ { name = "threadpoolctl", marker = "python_full_version < '3.11'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/98/c2/a7855e41c9d285dfe86dc50b250978105dce513d6e459ea66a6aeb0e1e0c/scikit_learn-1.7.2.tar.gz", hash = "sha256:20e9e49ecd130598f1ca38a1d85090e1a600147b9c02fa6f15d69cb53d968fda", size = 7193136, upload-time = "2025-09-09T08:21:29.075Z" }
wheels = [
@@ -2509,11 +2526,11 @@ resolution-markers = [
"(python_full_version == '3.11.*' and platform_machine != 'x86_64') or (python_full_version == '3.11.*' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "joblib" },
- { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
- { name = "scipy", version = "1.17.1", source = { registry = "https://pypi.org/simple" } },
- { name = "threadpoolctl" },
+ { name = "joblib", marker = "python_full_version >= '3.11'" },
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "scipy", version = "1.17.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
+ { name = "threadpoolctl", marker = "python_full_version >= '3.11'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/0e/d4/40988bf3b8e34feec1d0e6a051446b1f66225f8529b9309becaeef62b6c4/scikit_learn-1.8.0.tar.gz", hash = "sha256:9bccbb3b40e3de10351f8f5068e105d0f4083b1a65fa07b6634fbc401a6287fd", size = 7335585, upload-time = "2025-12-10T07:08:53.618Z" }
wheels = [
@@ -2552,8 +2569,8 @@ resolution-markers = [
"(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
]
sdist = { url = "https://files.pythonhosted.org/packages/0f/37/6964b830433e654ec7485e45a00fc9a27cf868d622838f6b6d9c5ec0d532/scipy-1.15.3.tar.gz", hash = "sha256:eae3cf522bc7df64b42cad3925c876e1b0b6c35c1337c93e12c0f366f55b0eaf", size = 59419214, upload-time = "2025-05-08T16:13:05.955Z" }
wheels = [
@@ -2616,8 +2633,8 @@ resolution-markers = [
"(python_full_version == '3.11.*' and platform_machine != 'x86_64') or (python_full_version == '3.11.*' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
]
sdist = { url = "https://files.pythonhosted.org/packages/7a/97/5a3609c4f8d58b039179648e62dd220f89864f56f7357f5d4f45c29eb2cc/scipy-1.17.1.tar.gz", hash = "sha256:95d8e012d8cb8816c226aef832200b1d45109ed4464303e997c5b13122b297c0", size = 30573822, upload-time = "2026-02-23T00:26:24.851Z" }
wheels = [
@@ -2744,8 +2761,8 @@ name = "standard-aifc"
version = "3.13.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "audioop-lts" },
- { name = "standard-chunk" },
+ { name = "audioop-lts", marker = "python_full_version >= '3.13'" },
+ { name = "standard-chunk", marker = "python_full_version >= '3.13'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/c4/53/6050dc3dde1671eb3db592c13b55a8005e5040131f7509cef0215212cb84/standard_aifc-3.13.0.tar.gz", hash = "sha256:64e249c7cb4b3daf2fdba4e95721f811bde8bdfc43ad9f936589b7bb2fae2e43", size = 15240, upload-time = "2024-10-30T16:01:31.772Z" }
wheels = [
@@ -2766,7 +2783,7 @@ name = "standard-sunau"
version = "3.13.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "audioop-lts" },
+ { name = "audioop-lts", marker = "python_full_version >= '3.13'" },
]
sdist = { url = "https://files.pythonhosted.org/packages/66/e3/ce8d38cb2d70e05ffeddc28bb09bad77cfef979eb0a299c9117f7ed4e6a9/standard_sunau-3.13.0.tar.gz", hash = "sha256:b319a1ac95a09a2378a8442f403c66f4fd4b36616d6df6ae82b8e536ee790908", size = 9368, upload-time = "2024-10-30T16:01:41.626Z" }
wheels = [
@@ -2791,6 +2808,7 @@ name = "stemdeck"
source = { editable = "." }
dependencies = [
{ name = "all-in-one-infer" },
+ { name = "asteroid-filterbanks" },
{ name = "audio-separator", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
{ name = "audioread", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
{ name = "beat-this" },
@@ -2825,6 +2843,7 @@ dev = [
[package.metadata]
requires-dist = [
{ name = "all-in-one-infer", specifier = ">=3.1,<4" },
+ { name = "asteroid-filterbanks", specifier = ">=0.4" },
{ name = "audio-separator", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'", specifier = ">=0.24" },
{ name = "audioread", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'", specifier = ">=3.0" },
{ name = "beat-this", specifier = ">=1.1" },
@@ -2932,13 +2951,13 @@ resolution-markers = [
"python_full_version < '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'",
]
dependencies = [
- { name = "filelock" },
- { name = "fsspec" },
- { name = "jinja2" },
- { name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "networkx", version = "3.6.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
- { name = "sympy" },
- { name = "typing-extensions" },
+ { name = "filelock", marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "fsspec", marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "jinja2", marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "networkx", version = "3.6.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "sympy", marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
+ { name = "typing-extensions", marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/3b/55/7192974ab13e5e5577f45d14ce70d42f5a9a686b4f57bbe8c9ab45c4a61a/torch-2.2.2-cp310-none-macosx_10_9_x86_64.whl", hash = "sha256:b2e2200b245bd9f263a0d41b6a2dab69c4aca635a01b30cca78064b0ef5b109e", size = 150788930, upload-time = "2024-03-27T21:08:09.98Z" },
@@ -2957,11 +2976,11 @@ resolution-markers = [
"(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "filelock" },
- { name = "fsspec" },
- { name = "jinja2" },
- { name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "networkx", version = "3.6.1", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
+ { name = "filelock", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "fsspec", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "jinja2", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "networkx", version = "3.4.2", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "networkx", version = "3.6.1", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
{ name = "nvidia-cublas-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
{ name = "nvidia-cuda-cupti-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
{ name = "nvidia-cuda-nvrtc-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
@@ -2975,10 +2994,10 @@ dependencies = [
{ name = "nvidia-nccl-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
{ name = "nvidia-nvjitlink-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
{ name = "nvidia-nvtx-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
- { name = "setuptools", marker = "python_full_version >= '3.12'" },
- { name = "sympy" },
+ { name = "setuptools", marker = "(python_full_version >= '3.12' and platform_machine != 'x86_64') or (python_full_version >= '3.12' and sys_platform != 'darwin')" },
+ { name = "sympy", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
{ name = "triton", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
- { name = "typing-extensions" },
+ { name = "typing-extensions", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/37/81/aa9ab58ec10264c1abe62c8b73f5086c3c558885d6beecebf699f0dbeaeb/torch-2.6.0-cp310-cp310-manylinux1_x86_64.whl", hash = "sha256:6860df13d9911ac158f4c44031609700e1eba07916fff62e21e6ffa0a9e01961", size = 766685561, upload-time = "2025-01-29T16:19:12.12Z" },
@@ -3009,7 +3028,7 @@ resolution-markers = [
"python_full_version < '3.11' and platform_machine == 'x86_64' and sys_platform == 'darwin'",
]
dependencies = [
- { name = "torch", version = "2.2.2", source = { registry = "https://pypi.org/simple" } },
+ { name = "torch", version = "2.2.2", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine == 'x86_64' and sys_platform == 'darwin'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/76/70/ca793994d37815070f6b53932b71822f66cfb3e197e6937426815998221e/torchaudio-2.2.2-cp310-cp310-macosx_10_13_x86_64.whl", hash = "sha256:b1d58201d108e85db3e35b84319f33884f61f327c38ead86913218c8c1acc3dd", size = 3398751, upload-time = "2024-03-27T21:12:31.998Z" },
@@ -3028,7 +3047,7 @@ resolution-markers = [
"(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')",
]
dependencies = [
- { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" } },
+ { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/38/aa/f634960ac094e3fc6869f5c214ccfa6f74da2b1a89cefac024f6c650a717/torchaudio-2.6.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:0eda1cd876f44fc014dc04aa680db2fa355a83df5d834398db6dd5f5cd911f4c", size = 1808471, upload-time = "2025-01-29T16:29:43.783Z" },
@@ -3054,10 +3073,10 @@ name = "torchvision"
version = "0.21.0"
source = { registry = "https://pypi.org/simple" }
dependencies = [
- { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.11'" },
- { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.11'" },
- { name = "pillow" },
- { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" } },
+ { name = "numpy", version = "2.2.6", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version < '3.11' and platform_machine != 'x86_64') or (python_full_version < '3.11' and sys_platform != 'darwin')" },
+ { name = "numpy", version = "2.4.4", source = { registry = "https://pypi.org/simple" }, marker = "(python_full_version >= '3.11' and platform_machine != 'x86_64') or (python_full_version >= '3.11' and sys_platform != 'darwin')" },
+ { name = "pillow", marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
+ { name = "torch", version = "2.6.0", source = { registry = "https://pypi.org/simple" }, marker = "platform_machine != 'x86_64' or sys_platform != 'darwin'" },
]
wheels = [
{ url = "https://files.pythonhosted.org/packages/a9/20/72eb0b5b08fa293f20fc41c374e37cf899f0033076f0144d2cdc48f9faee/torchvision-0.21.0-1-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:5568c5a1ff1b2ec33127b629403adb530fab81378d9018ca4ed6508293f76e2b", size = 2327643, upload-time = "2025-03-18T17:25:51.165Z" },