Feature: Add SenseVoice as STT engine — 10x faster than Whisper, 50+ languages

[LauraGPT]

Hi! Voicebox is an incredible project. I noticed the roadmap mentions adding more STT engines (Parakeet v3, Qwen3-ASR).

I'd like to suggest SenseVoice as another engine option. It's a non-autoregressive ASR model that's particularly well-suited for dictation/voice input use cases:

Why SenseVoice fits Voicebox well:

• ~10x faster than Whisper Large — 50ms for 10s audio on GPU, making dictation feel instant
• Non-autoregressive: Single forward pass, no beam search — extremely predictable latency
• 50+ languages including Chinese, English, Japanese, Korean, French, German, Spanish, etc.
• Built-in emotion detection: Could enrich Captures with emotional context
• 234M parameters — smaller than Whisper Turbo, runs great on consumer hardware
• ONNX/MLX compatible via Sherpa-ONNX for Apple Silicon optimization

Integration via FunASR (Python/PyTorch):

from funasr import AutoModel

model = AutoModel(model="iic/SenseVoiceSmall")
result = model.generate(input=audio_chunk)
text = result[0]["text"]

Or for Apple Silicon via Sherpa-ONNX (C++/Swift):
SenseVoice models are available in ONNX format and run natively on macOS/iOS via Sherpa-ONNX, which could integrate well with Voicebox's MLX architecture.

OpenAI-compatible server mode:

pip install funasr
funasr-server --device cuda  # or --device cpu for Mac
# Same /v1/audio/transcriptions endpoint

For the push-to-talk dictation workflow, the speed difference between Whisper and SenseVoice is very noticeable — text appears almost instantly after releasing the hotkey. Happy to help with integration!

• SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8K+ stars)
• FunASR toolkit: https://github.com/modelscope/FunASR (16K+ stars)
• Sherpa-ONNX (cross-platform): https://github.com/k2-fsa/sherpa-onnx (5K+ stars)