Feature: Add SenseVoice as alternative to Whisper — 10× faster, 234M params

[LauraGPT]

Note

License and capability clarification (2026-07-14): FunASR is a toolkit, not a single checkpoint. The FunASR and SenseVoice repository source code is MIT; model weights follow each model card. SenseVoiceSmall supports Chinese, Cantonese, English, Japanese, and Korean, and its weights use the linked FunASR Model Open Source License Agreement. Fun-ASR-Nano-2512 is Apache-2.0. Language coverage, punctuation, and performance depend on the selected model and runtime configuration.

Feature Request

Amical currently uses Whisper for speech-to-text. SenseVoice would be a great addition as a faster, lighter alternative:

Comparison

	Whisper (Small)	SenseVoice Small
Architecture	Autoregressive	Non-autoregressive
Parameters	244M	234M
Speed	~30× realtime	~170× realtime
Languages	99	50+
Extras	—	Emotion detection, audio events

Why it matters for Amical:

• 10× faster inference → more responsive dictation
• Same model size as Whisper Small but dramatically faster
• Non-autoregressive → no hallucination/repetition issues that affect Whisper
• Privacy-first: fully local, same as Whisper

Easy to integrate

pip install funasr

from funasr import AutoModel
model = AutoModel(model="iic/SenseVoiceSmall")
result = model.generate(input="audio.wav")
text = result[0]["text"]

Or via ONNX (same format as Whisper ONNX):

• sherpa-onnx provides cross-platform C++ inference with Python/Swift bindings
• ~400MB ONNX model download

References

• https://github.com/FunAudioLLM/SenseVoice (8.3K ⭐)
• https://github.com/modelscope/FunASR (16.6K ⭐)