add sensevoice stt model

[mm65x]

Context

SenseVoice Small (https://github.com/FunAudioLLM/SenseVoice) from Alibaba DAMO Academy. does transcription for 50+ languages (strongest on zh/en/ja/ko/yue) plus language id, emotion, and audio event detection. 234M params, runs at about 70ms for 10s of audio on apple silicon.

Description

adds sensevoice to the stt pipeline. the model uses SANM (self-attention with FSMN memory) layers, which are a bit unusual - each attention layer has a parallel depthwise conv branch that acts as a memory mechanism.

the upstream repo only distributes pytorch .pt weights with a yaml config, so this includes a conversion script that produces safetensors + a standard config.json. once there's a converted model on mlx-community/huggingface this won't be needed for most users.

Changes in the codebase

• mlx_audio/stt/models/sensevoice/sensevoice.py - model implementation (frontend, encoder, ctc head, generate)
• mlx_audio/stt/models/sensevoice/config.py - config dataclasses
• mlx_audio/stt/models/sensevoice/convert.py - pt → safetensors converter
• mlx_audio/stt/models/sensevoice/README.md - setup + usage
• mlx_audio/stt/models/sensevoice/tests/test_sensevoice.py - 16 tests
• mlx_audio/stt/models/__init__.py, mlx_audio/stt/utils.py - registration

Changes outside the codebase

none.

Additional information

• verified component by component against the FunASR and SenseVoice reference repos
• tested end-to-end on en, zh, ja, ko, yue - transcriptions, language detection, emotion, and event output all correct
• users need to run the conversion script for now (instructions in the README), could upload a pre-converted model to mlx-community later

Checklist

• Tests added/updated
• Documentation updated
• Issue referenced - STT roadmap

[Blaizzy]

LGTM!

Just have one nit before we merge 👌🏽

[Blaizzy]

Could you move this into the correct tests folder?

[lucasnewman]

@mm65x Would you mind uploading to converted model to mlx-community to make it easy for folks to use?

[mm65x]

moved to mlx_audio/stt/tests/ in latest commit. let me know if that's the right spot!

i'll also try to get the converted weights up on mlx-community later today.

[mm65x]

@lucasnewman i don't have write access to the mlx-community org on huggingface, so i can't upload it there directly. i included the conversion script in this PR so anyone with access should be able to run it and upload the weights! let me know if you want me to upload it to my personal HF instead.

[lucasnewman]

@lucasnewman i don't have write access to the mlx-community org on huggingface, so i can't upload it there directly. i included the conversion script in this PR so anyone with access should be able to run it and upload the weights! let me know if you want me to upload it to my personal HF instead.

You can just add yourself to the community to upload models, but I can do it for you in a bit if needed.

[lucasnewman]

@mm65x I uploaded the model here: https://huggingface.co/mlx-community/SenseVoiceSmall

Can you update the README to use mlx-community/SenseVoiceSmall and remove the conversion section? Then we can merge.

[mm65x]

@lucasnewman oh nice, i see you uploaded it to mlx-community/SenseVoiceSmall! i've just removed the convert.py script from this PR and updated the README to point directly to the huggingface repo. let me know if there's anything else needed here.

[lucasnewman]

🚀