Feature: Add SenseVoice support via Sherpa-ONNX for on-device ASR

[LauraGPT]

Motivation

ElatoAI runs realtime voice AI on ESP32. SenseVoice (8K+ stars) is available through Sherpa-ONNX which already supports embedded/IoT platforms including ESP32-compatible inference.

Why SenseVoice for ElatoAI

• Non-autoregressive: Single forward pass, minimal compute per chunk
• SenseVoice-Small (234M): 50+ languages with auto detection
• ONNX format: Runs via Sherpa-ONNX on embedded devices
• Built-in VAD: No separate voice activity detection needed
• Emotion detection: Detect user emotions from speech — useful for companion AI

Integration via Sherpa-ONNX

Sherpa-ONNX provides C/C++ API suitable for embedded:

// C API for embedded devices
SherpaOnnxOfflineRecognizer *recognizer = 
    SherpaOnnxCreateOfflineRecognizer(&config);
// Process audio frames → get text

For server-side processing (when ESP32 sends audio to a server):

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

References

• FunASR: https://github.com/modelscope/FunASR
• Sherpa-ONNX (C/C++): https://github.com/k2-fsa/sherpa-onnx