Faster + more accurate alternative: FunASR SenseVoice (234MB, 170x realtime)
#50
by langgz - opened
If you're using whisper-small for its smaller size and speed, FunASR SenseVoice might interest you:
| whisper-small | FunASR SenseVoice | |
|---|---|---|
| Size | 244MB | 234MB |
| GPU Speed | ~37x realtime | 170x realtime |
| CPU Speed | slow | 17x realtime |
| Languages | 57 | 5 (zh/en/ja/ko/yue) |
| Speaker ID | ❌ | ✅ built-in |
| Emotion | ❌ | ✅ |
Similar model size, but 4-5x faster and includes speaker diarization + emotion detection.
pip install funasr
from funasr import AutoModel
model = AutoModel(model="FunAudioLLM/SenseVoiceSmall", hub="hf", vad_model="funasr/fsmn-vad", device="cuda")
result = model.generate(input="audio.wav")
GitHub: https://github.com/modelscope/FunASR (16K+ ⭐)