Feature Request: Add SenseVoice/Paraformer as faster ASR option

[LauraGPT]

Feature Request

whisper-diarization is excellent for ASR + speaker diarization. Suggesting SenseVoice / Paraformer as alternative ASR backends — they're much faster and FunASR includes its own diarization model.

Why SenseVoice/Paraformer?

• 5-10x faster than Whisper — non-autoregressive, dramatically reduces processing time for long recordings
• Built-in speaker diarization — FunASR includes cam++ (7.2M params) for speaker embedding, no separate pyannote needed
• Built-in VAD — FSMN-VAD with accurate timestamps
• Built-in punctuation — automatic punctuation restoration
• 50+ languages — SenseVoice handles multilingual content

Complete pipeline comparison

Current: Whisper + pyannote → post-processing alignment
Alternative: FunASR (SenseVoice + FSMN-VAD + cam++ + CT-Punc) → all-in-one

from funasr import AutoModel

model = AutoModel(
    model="iic/SenseVoiceSmall",
    vad_model="fsmn-vad",
    spk_model="cam++",
    punc_model="ct-punc",
)
result = model.generate(input="meeting.wav")
# Returns: text + timestamps + speaker labels + punctuation

Speed benefit for long recordings

For a 1-hour recording:

• Whisper large-v3: ~10-15 minutes processing
• SenseVoice: ~2-3 minutes processing (5x faster)

Install: pip install funasr

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