Feature: Add FunASR/SenseVoice as alternative ASR backend

## Motivation WhisperLive provides near-realtime transcription using Whisper. [FunASR](https://github.com/modelscope/FunASR) (16K+ stars) offers a faster alternative backend, especially beneficial for live transcription where latency matters. ## Why FunASR for Live Transcription - **Non-autoregressive**: SenseVoice processes audio in a single forward pass — no iterative token generation, lower latency per chunk - **5x faster** than Whisper large-v3 at comparable accuracy - **Smaller model** (234M params vs 1.5B), less GPU memory, faster load - **Built-in streaming ASR**: FunASR includes a WebSocket-based streaming server for real-time recognition - **Built-in VAD**: FSMN-VAD (5.2M params) — no separate Silero VAD needed - **50+ languages** with auto detection ## FunASR Streaming Server FunASR already has a production-grade streaming ASR server: ```bash pip install funasr vllm funasr-server --device cuda # OpenAI-compatible API at :8000 # Real-time WebSocket streaming # ws://localhost:8000/ws (streaming ASR endpoint) ``` Or via Python API: ```python from funasr import AutoModel model = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad") result = model.generate(input="audio_chunk.wav") ``` ## References - GitHub: https://github.com/modelscope/FunASR - Streaming: WebSocket-based realtime ASR - pip: `pip install funasr`