Feature: Add FunASR/SenseVoice as STT extension

[LauraGPT]

Motivation

TEN Framework is a powerful conversational voice AI framework. FunASR (16K+ stars) could be a great addition as an STT extension, particularly strong for Chinese and multilingual scenarios.

Why FunASR

• OpenAI-compatible API: /v1/audio/transcriptions — easy to integrate as an STT service
• 50+ languages with auto language detection
• Non-autoregressive: SenseVoice processes in single forward pass — lower latency for conversational AI
• Built-in VAD: FSMN-VAD, no external dependency needed
• Built-in speaker diarization: cam++ model for multi-speaker scenarios
• Emotion detection: SenseVoice detects speech emotions — useful for conversational AI context
• Streaming ASR: WebSocket-based real-time recognition
• Free & self-hosted: No API costs

Integration as TEN Extension

# Start FunASR server
pip install funasr vllm
funasr-server --device cuda  # OpenAI-compatible at :8000

Since TEN already supports OpenAI-compatible STT services, FunASR can be used as a drop-in backend. A dedicated TEN extension could provide:

• Real-time streaming via WebSocket
• Emotion metadata for agent reactions
• Speaker identification in multi-party conversations

References

• GitHub: https://github.com/modelscope/FunASR
• Streaming ASR: WebSocket endpoint
• Models: SenseVoice (50+ langs), Paraformer (zh/en high accuracy)