[Showcase] recording-inbox — Feishu Drive inbox → local FunASR/SenseVoice → structured meeting minutes

[chenmozhe008]

Summary

recording-inbox is an open-source, self-hosted workflow: recordings uploaded from iPhone/Android/PC to a Feishu (Lark) Drive folder are picked up by a Mac/Windows machine, transcribed locally with FunASR (SenseVoiceSmall), turned into AI meeting minutes plus a full timestamped transcript, and published back to Feishu as two documents with a single completion notification.

Scenario

• Local transcription demo
• Batch transcription / archive processing
• Other: personal meeting/interview recording automation (Feishu-centric, idempotent resumable pipeline)

Deployment details

• FunASR version: installed unpinned from PyPI (pip install funasr modelscope torch torchaudio ..., see macOS setup)
• Model(s): iic/SenseVoiceSmall (primary backend), whisper.cpp as failure fallback
• Device (cuda, cpu, mps): cpu (Apple Silicon Mac); also verified with CUDA explicitly disabled
• GPU / CPU: Apple Silicon (M-series), CPU-only inference
• Runtime path: Python API
• Audio language/domain: Chinese meetings, interviews and lectures (SenseVoice multilingual support available)
• Approximate audio duration or traffic: personal daily use; recordings from tens of seconds up to multi-hour, processed by a background scheduled task

Results

- Fresh-environment deployment: GitHub clone -> venv -> FunASR install (incl. first 1.2 GB model
  download) -> full pipeline finished for a 30 s recording in 2 min 06 s, no fallback triggered.
- Independent verification by a FunASR maintainer at commit 389416a: a real 45.1 s audio file
  transcribed successfully with funasr-sensevoice in both the default environment and a
  CPU-only (CUDA disabled) environment, with identical transcript hashes across runs.
- 25 unit / mocked end-to-end tests pass; Linux, macOS and Windows GitHub Actions are green.
  (Public CI is a mocked pipeline and does not download models; real inference was verified
  separately as above.)

Links or screenshots

• Repository: https://github.com/chenmozhe008/recording-inbox
• Project page: https://chenmozhe008.github.io/recording-inbox/
• Maintainer's independent verification record: [开源推荐] 录音上传后自动生成飞书智能纪要 521xueweihan/HelloGitHub#3431 (comment)
• Maintenance status: community-maintained, actively used daily by the author; scope intentionally limited to the "upload -> local ASR -> structured minutes -> Feishu archive/notify" loop

What was easy?

• SenseVoiceSmall works out of the box on CPU — no GPU required for a personal daily workload.
• Automatic model download via modelscope made first-run setup a single pip install.
• Built-in timestamps map directly to our per-segment transcript format.

What should be improved?

• The first-run 1.2 GB model download is the biggest install hurdle for non-developer users on slow connections; an official slimmer variant or resumable-download guidance would help.

[LauraGPT]

Thanks for the detailed, reproducible showcase. The local CPU validation, multi-hour workflow scope, and public setup guide meet the community-integration evidence bar.

I added recording-inbox to both the English and Chinese community integration pages in #3295, now merged on main. This gives the Feishu/Lark -> local SenseVoiceSmall -> structured minutes workflow a stable discovery entry.

One small upstream hardening item remains worth considering independently: pin a tested funasr version in the setup docs so future dependency changes do not silently alter deployments.

[chenmozhe008]

Thank you for the integration and for the careful review.

We have implemented the suggested deployment hardening in recording-inbox#21:

• new installs pin the inference-tested core baseline to funasr==1.3.10 and modelscope==1.37.1;
• macOS and Windows use explicit platform requirement files;
• setup_check.py reports the installed FunASR, ModelScope, and PyTorch versions and warns (without breaking an existing working deployment) when the core baseline differs;
• the setup guides now explain first-run model downloads and persistent local caching;
• the unused torchaudio dependency was removed after dependency resolution showed that current macOS wheels can select mismatched Torch/Torchaudio releases.

We intentionally leave PyTorch unpinned so pip can select a wheel compatible with the user's OS, CPU, and Python version, while making the selected version visible in diagnostics.

Thanks again for adding recording-inbox to both community integration pages.