Feature: Add SenseVoice for lower-latency real-time transcription

[LauraGPT]

Hi! Impressive AI meeting copilot with echo cancellation.

For real-time meeting transcription, SenseVoice could significantly reduce ASR latency:

Key advantages for a meeting copilot

1. Non-autoregressive — single forward pass gives full transcription (no sequential token generation)
2. ~50ms latency on GPU — ideal for real-time copilot scenarios
3. 234M params / ~1GB VRAM — leaves room for your LLM on the same GPU
4. Built-in features: VAD, speaker diarization (cam++), emotion detection
5. OpenAI-compatible API — if you already use Whisper API, it's a drop-in

Quick start

from funasr import AutoModel

model = AutoModel(
    model="iic/SenseVoiceSmall",
    vad_model="fsmn-vad",
    spk_model="cam++",
)
result = model.generate(input=audio_chunk)

Or start a server:

pip install funasr
funasr-server --device cuda  # localhost:8000, OpenAI-compatible

Links

• SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8.3K stars)
• FunASR: https://github.com/modelscope/FunASR (16.7K stars)

[berasamiran1204-eng]

Thank you

…

On Wed, 3 Jun, 2026, 4:11 pm Ruo Tian, ***@***.***> wrote: *fatumax* left a comment (Laxcorp-Research/project-raven#1) <#1 (comment)> @LauraGPT <https://github.com/LauraGPT> 我最初给这个项目点 Star 的时候是真心觉得不错，思路新颖，想法有趣。 但这段时间越来越觉得不对劲了。Star 数一直在涨，项目本身却停滞不前：代码没什么更新，Issues 没人搭理，文档该缺的还缺。一个号称有这么多人关注的项目，怎么可能连最起码的维护都做不到？ 后来我认真查看了近期打星的那些账号，心里仅存的疑虑变成了确信。那些账号的特征如出一辙，不可能是真实的独立用户。 我不愤怒，只是感到深深的失落。我当时那一颗出于真心的 Star，如今被淹没在一堆虚假数字当中，变得一文不值。 所以我取消了 Star。并非出于报复，而是不愿意让自己的那份认可，沦为一场骗局的注脚。 衷心希望你们终有一天能凭真正的实力和诚意，赢回社区的尊重。 — Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=CA6Z4IWU5I5ERW5V2L2XKRL4576G3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINRRGE2DQMBTGE3KM4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#issuecomment-4611480316>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/CA6Z4IVHK2ZOPEK5LKA2GJ34576G3AVCNFSM6AAAAACZUMEUAWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DMMJRGQ4DAMZRGY> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/CA6Z4IVPH33CXF74XHGDXV34576G3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINRRGE2DQMBTGE3KM4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSVGM33PORSXEX3JN5ZQ> and Android <https://github.com/notifications/mobile/android/CA6Z4IRMN4BRVFJRHSMXZUD4576G3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINRRGE2DQMBTGE3KM4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSXGM33PORSXEX3BNZSHE33JMQ>. Download it today! You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[claxman]

Thanks for the detailed write-up @LauraGPT — the SenseVoice numbers (non-autoregressive, ~50ms on GPU, built-in VAD/diarization) are genuinely compelling for real-time transcription.

One architectural note on fit: Raven is an Electron desktop client that streams audio to a hosted streaming endpoint (Deepgram over WebSocket today) rather than running model inference locally. Taking a hard dependency on funasr would mean bundling/operating a Python + GPU inference stack, which is a much bigger lift for an end-user desktop app and not something we can ship by default.

Rather than hard-wiring any single engine, the right move here is a pluggable transcription-provider interface — a small contract (start / stream PCM / is_final results / stop) with Deepgram as the reference implementation, so alternative providers can be wired in without touching the core pipeline. That's the direction we're taking.

Because SenseVoice/FunASR exposes an OpenAI-compatible API, it would slot into that abstraction cleanly once a SenseVoice server is reachable (self-hosted or hosted) — no special-casing required. We'll keep this issue open to track that provider interface landing, and we'd welcome a community provider impl for SenseVoice on top of it.