[sw005320]

All five course notebooks run, each on its own workflow, every Sunday.

The PR started as four of five with the fifth stopped by a licence. It is five
now, and the last one needs a decision from a reader rather than from me — the
second bullet under Samples below.

neural_codec — three missing packages, not a broken scorer

I said VERSA's scorer "exits without writing its output". It does not. It
refuses the config and says exactly why:

scorer.py: error: Invalid score configuration:
- score_config[0] (signal_metric): missing optional dependencies: ci_sdr, fast_bss_eval, mir_eval.
  Install them or remove this metric from the config

signal_metric wants VERSA's optional extras. Nothing to remove — the notebook
asks for the three packages, and the scorer produces its sdr, sir, sar,
si_snr and ci_sdr as intended. My earlier run only "worked" because that
environment happened to have fast_bss_eval from espnet[enh].

speech_translation — nothing was wrong with it

It was "not tried" because the numpy==1.23.5 / tensorflow==2.12.0 pins made
trying pointless. With those gone it runs: SimulEval, ParallelWaveGAN,
pysndfile and all.

speech_enhancement — samples the repository cannot carry

Both of its recordings came from Google Drive copies of licensed corpora.
Neither can be committed here, so each is now something the notebook fetches
or builds itself.

CHiME-4. The real noisy sample is ped.wav from the challenge's own
data page, which
publishes a few recordings to listen to. The corpus is LDC2017S24 and built
on WSJ0. The page's sample is one channel where the corpus has six; the
enhancement in the notebook is single-channel either way, so the demonstration
is unchanged.

wsj0-2mix → Libri2Mix. The separation half used a mixture from wsj0-2mix,
which is built on WSJ0 and came from Drive. Libri2Mix is built on LibriSpeech,
which is CC BY 4.0, so it uses the real thing instead: mixture
7729-102255-0031_2094-142345-0028 of the Libri2Mix test set, built in the
notebook from its two sources the way LibriMix builds it — the gains the
official metadata gives for that pair, resampled to 8 kHz, truncated to the
shorter source (min mode). The model moved with the data, and so did the ASR
after it.

Having both sources is what makes it worth doing. The notebook now scores the
separation instead of only playing it:

speaker 1:   3.24 dB in the mixture ->  16.21 dB separated
speaker 2:  -3.28 dB in the mixture ->  12.88 dB separated

Mix Speech to Text:  HE CONTINUED ON HIS GROUND AND SIR HOWDY HUNDRED HE OF WET COLLARED TO HIS LATE ERRAND CARRYING TWO AROUND
Separated Speech 1:  HE CONTINUED HIS PRETENDED SEARCH AND TO GIVE COLOR TO HIS ERRAND MADE TO ARREST
Separated Speech 2:  YOU KNOW NO NO TOTTY HAD AT HER FEET WET SAID MISSUS POYSER CARRYING AWAY HER IRON

Two and ten characters from the references, against a mixture that comes out
as nonsense.

Five models in the organisation could not be loaded

The Libri2Mix model did not load at all:

size mismatch for separator.tcn.network.3.weight: copying a param with shape
torch.Size([1024, 128, 1]) from checkpoint, the shape in current model is
torch.Size([1])

espnet's TCN gained a non-linearity before the mask convolution in May 2023
(02b9cc722c9). The layer kept the old layout as its default, but
TCNSeparator — what a config builds — defaults to prelu, so every TCN
checkpoint older than that stopped loading. Five models in the espnet
organisation were affected, and their configs now record the layout their
weights were trained with; each was loaded and run before its config was
uploaded. (A sixth, Wangyou_Zhang_chime4_enh_train_enh_conv_tasnet_raw,
already carried the key — someone had hit this before.)

Two more, YushiUeda_librimix_diar_enh_2_3_spk and its _lmf sibling, are
broken by something else: they were trained by a joint diarization and
enhancement task that current espnet does not have, and they fail before the
separator is built. They are untouched.

I did not change espnet's default. No egs2 config sets
pre_mask_nonlinear, so flipping it would change what a new training run
builds, which is a maintainer's call rather than mine.

The tables are checked now, not written

check_layout.py held only Demos/. The same rules hold for the course too: a
workflow that exists and names the notebook, a badge in the notebook and in
every README that lists it, a pinned release, and no dead notebook links. The
pre-commit trigger covers the READMEs and the workflows rather than only the
notebooks — a README edited on its own is exactly how the course table came to
claim four of five on a day when all five ran.

Also here: device="cuda" if ... else "cpu" if ... else "cpu" in two notebooks,
left by an earlier edit of mine, and two __pycache__ directories that were
committed and are now ignored.

Checked

• speech_enhancement.ipynb run end to end locally, RESULT ok, with both
  substituted samples and no gdown.
• The five per-notebook workflows run on this PR — that is the check that
  matters, and the badges follow it.
• tools/check_layout.py clean, tests/ 23 passed.

🤖 Generated with Claude Code

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request