SenseVoice realtime transcription

ws: — vad: — mic: —
server
level idle

Mic audio is captured in the browser, downsampled to 16 kHz mono, base64-PCM16 encoded, and streamed to /v1/realtime?intent=transcription. FSMN-VAD end-points each sentence: provisional text updates in place; when a sentence is committed it moves onto its own line. Serve the page yourself, or run the server with --web runtime/llama.cpp/sensevoice-server/webui and open http://127.0.0.1:8040/.