feat(funasr): add FunASRTranscriber component

[SyedShahmeerAli12]

Related Issues

• fixes Feature Request: Add FunASR Speech-to-Text component for audio document processing #3375

Proposed Changes:

Adds a FunASRTranscriber component that transcribes audio files to Haystack Document objects using FunASR an open-source, self-hosted speech recognition toolkit from Alibaba DAMO Academy (no API key required).

• Accepts str, Path, and ByteStream sources; temp files for ByteStream are always cleaned up
• Default model iic/SenseVoiceSmall supports 50+ languages, 5-10x faster than Whisper
• Returns one Document per source; transcript in content, optional timestamps and speakers in metadata
• warm_up() for lazy model loading inside Haystack pipelines
• generate_kwargs pass-through for model-specific options (use_itn, merge_vad, language, hotword, etc.)
• Fully serialisable via to_dict / from_dict

How did you test it?

25 unit tests covering: init defaults/custom, serialisation round-trip, warm_up idempotency, single/multi-file transcription, VAD segment merging, metadata (single + list), timestamps, speakers, empty result, error skipping, generate_kwargs forwarding, ByteStream handling, and temp-file cleanup (including on error).

hatch run fmt-check   # clean
hatch run test:types  # clean
hatch run test:unit   # 25/25 passed

Integration tests (marked @pytest.mark.integration) are gated behind FUNASR_INTEGRATION_TESTS=1 to avoid downloading large model weights in CI.

Notes for the reviewer

• warm_up() uses a lazy from funasr import AutoModel inside the method body so that importing the package does not require torch at import time. Unit tests mock this via patch.dict("sys.modules", {"funasr": mock_funasr}).
• PyPy classifier was omitted FunASR depends on C extensions that are CPython-only.

Checklist

• I have read the contributors guidelines and the code of conduct
• I have updated the related issue with new insights and changes
• I added unit tests and updated the docstrings
• I've used one of the conventional commit types for my PR title: feat:

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	torch@​2.12.0
	funasr@​1.3.9
	torchaudio@​2.11.0

View full report

[davidsbatista]

There are unrelated files/changes in this PR to the issue #3375 - please review your PR

[SyedShahmeerAli12]

There are unrelated files/changes in this PR to the issue #3375 - please review your PR

yess i saw i am on it

[SyedShahmeerAli12]

Unit tests use mocked AutoModel real transcription was not tested locally (requires torch). To verify pip install funasr torch torchaudio FUNASR_INTEGRATION_TESTS =1 hatch run test:integration

[anakin87]

Thank you!

I left some suggestions

[anakin87]

Let's have this test running on the CI. It's ok if it takes 2-3 minutes.
Let's also use a short real audio file: one of these from Haystack would be OK
https://github.com/deepset-ai/haystack/tree/main/test/test_files/audio

[SyedShahmeerAli12]

The CI workflow already had the integration test step (pytest -m "integration"). It was just always skipping because of an env var check at the top of the test. Removed that check the test now runs for real, downloads answer.wav from Haystack's test files, and uses iic/SenseVoiceSmall to transcribe it.

[SyedShahmeerAli12]

removed LazyImport in favour of a plain top-level import since funasr is a required dependency; moved torch and torchaudio to core project dependencies; updated the model parameter link to the FunASR model selection page; and fixed the integration test to use answer.wav from Haystack's test files without any env var gate.

[anakin87]

A few final comments

[anakin87]

Pushed some final changes, including categorizing this component as audio (consistent with https://docs.haystack.deepset.ai/docs/remotewhispertranscriber)

Thank you!