Add FunASR recognizer (local multilingual ASR: SenseVoice/Paraformer)

[LauraGPT]

Adds a FunASR recognizer: Recognizer.recognize_funasr(audio_data, model="iic/SenseVoiceSmall", ...).

FunASR is an open-source speech toolkit (SenseVoice, Paraformer, Fun-ASR-Nano) that runs fully locally, no API key, with strong multilingual accuracy — particularly Chinese, Cantonese, Japanese, Korean. SenseVoice (the default) is multilingual with auto language detection.

Implementation mirrors the existing local recognizers (e.g. vosk): a recognize() in speech_recognition/recognizers/funasr.py that converts AudioData to 16 kHz/16-bit PCM via get_raw_data(convert_rate=16000, convert_width=2), runs the model, and returns cleaned text. The model is cached across calls; funasr is an optional extra (pip install "SpeechRecognition[funasr]").

Verified: the AudioData.get_raw_data(16000, 2) -> int16 -> float32 conversion was tested on a 16 kHz sample, and the FunASR/SenseVoice transcription core (PCM -> text) was verified end-to-end (produces the expected transcript with tags stripped).

import speech_recognition as sr
r = sr.Recognizer()
with sr.AudioFile("chinese.wav") as src:
    audio = r.record(src)
print(r.recognize_funasr(audio))  # local, multilingual

[LauraGPT]

Added lightweight FunASR recognizer tests in e38853a.

What the tests cover without downloading a real model:

• missing optional funasr dependency raises SetupError;
• AudioData.get_raw_data(convert_rate=16000, convert_width=2) is used before converting PCM to float32 samples;
• AutoModel is cached by (model, device) across calls;
• unknown languages fall back to FunASR/SenseVoice auto language detection;
• Recognizer.recognize_funasr is wired to the FunASR recognizer.

Local verification:

python -m pytest tests/recognizers/test_funasr.py -q
# 4 passed, 2 warnings in 0.09s

python -m pytest tests/test_recognition.py tests/recognizers/test_funasr.py -q
# 5 passed, 6 skipped, 2 warnings in 0.07s

python -m py_compile speech_recognition/recognizers/funasr.py tests/recognizers/test_funasr.py speech_recognition/__init__.py
git diff --check

The warnings are the existing Python 3.12 aifc / audioop deprecation warnings from package import.

[LauraGPT]

CI status note: the current unstable state is not a test failure from this diff. The four GitHub Actions runs on e38853a7 (Type check, Static analysis, Ensure RST is well-formed, and Unit tests) are all in action_required, which means the fork PR workflows are waiting for maintainer approval to run.

Local validation already covered the new recognizer path without downloading a real FunASR model, as noted above. Once the workflows are approved, I can follow up on any real failures if they appear.

[LauraGPT]

Added a small docs/lint follow-up in 2dcc004.

Changes:

• list FunASR in the README engine support list;
• add the funasr optional dependency / SpeechRecognition[funasr] install note;
• add recognize_funasr(...) to reference/library-reference.rst;
• remove an unused patch import from tests/recognizers/test_funasr.py that flake8 would flag.

Verification run locally on the PR branch:

python3 -m pytest tests/recognizers/test_funasr.py -q
# 4 passed, 2 warnings

python3 -m pytest tests/test_recognition.py tests/recognizers/test_funasr.py -q
# 5 passed, 6 skipped, 2 warnings

python3 -m py_compile speech_recognition/recognizers/funasr.py tests/recognizers/test_funasr.py speech_recognition/__init__.py

rstcheck README.rst CONTRIBUTING.rst
# Success! No issues detected.
rstcheck --ignore-directives autofunction reference/*.rst
# Success! No issues detected.

flake8 --ignore=E501,E701,W503 --extend-exclude .venv,venv,build --doctests speech_recognition/recognizers/funasr.py tests/recognizers/test_funasr.py speech_recognition/__init__.py
# passed

Note: this PR's GitHub Actions are still waiting for maintainer approval to run on the fork, so I cannot inspect CI results yet. Once approved, I can follow up on any real failures.

[LauraGPT]

Fresh FunASR-side status recheck (2026-07-07 UTC):

• PR is open, non-draft, and mergeable at head 2dcc004a; mergeable_state=unstable is not from a failing check log.
• GitHub reports 4 check suites on this head, all completed/action_required, with 0 check-runs/logs. This is the fork-workflow approval gate for Type check, Static analysis, Ensure RST is well-formed, and Unit tests, not a diff failure.
• Thread-aware review audit shows 0 current unresolved review threads.
• Fresh local validation on the PR branch:
  • git diff --check -- . ':(exclude)README.rst' passed. (README.rst is excluded here because upstream carries CRLF/trailing-space-sensitive lines; the project CI validates RST separately.)
  • python3 -m py_compile speech_recognition/recognizers/funasr.py tests/recognizers/test_funasr.py passed.
  • python3 -m pytest tests/recognizers/test_funasr.py -q -> 4 passed.
  • README, pyproject.toml optional extra, and library reference entries for FunASR / recognize_funasr are present.

Remaining action is maintainer approval of the fork workflows so the repo CI can run on GitHub.

[LauraGPT]

@ftnext, could you approve the four fork workflow suites and review this local FunASR recognizer when convenient?

Fresh verification on exact head 2dcc004a against current master:

• branch is clean, mergeable, and 0 commits behind;
• isolated Python 3.12 .[dev] environment: full pytest 36 passed, 24 skipped;
• focused FunASR tests: 4 passed;
• mypy: no issues in 32 source files;
• flake8, both RST checks, Python compilation, and the CRLF-aware diff check pass;
• real CPU smoke with FunASR 1.3.14 and official iic/SenseVoiceSmall exercised AudioFile -> AudioData -> Recognizer.recognize_funasr, returned 开放时间早上9点至下午5点。, stripped rich tags, and populated exactly one model-cache entry.

The four GitHub check suites still have action_required with zero check runs, so there is no CI log or code failure for me to address until a maintainer approves them.