RapidAI/RapidSpeech-wasm

RapidSpeech.cpp — Browser Demo

Three-tab WebAssembly demo for RapidSpeech.cpp:

• —ASR Offline — upload a WAV (or record one) → optional neural VAD → transcript with timestamps.
• —ASR Online — live mic → neural VAD → streaming transcripts, with optional two-pass (CTC → LLM rescore).
• —TTS Offline — text → speech via OmniVoice / OpenVoice2 (with voice-cloning).

Everything runs locally in your browser via WebAssembly + WebGPU + pthreads. After the WASM module loads, audio and model bytes never leave the page.

Why Docker SDK?

The demo needs cross-origin isolation (COOP+COEP) so SharedArrayBuffer and pthreads are available. HuggingFace's static SDK does not reliably set those headers, so this Space uses the Docker SDK with a tiny Python server (serve.py) that emits the right headers explicitly.

Models

Pick any GGUF from RapidAI/RapidSpeech and paste the resolve URL into the demo. Recommended starters:

• —ASR: funasr-nano-q4_k_m.gguf (~600 MB)
• —VAD: silero-vad-v6.gguf (~2 MB)
• —TTS: omnivoice-q4_k_m.gguf or openvoice2-base-en.gguf

Browser requirements

• —Chromium 113+ for WebGPU (Chrome / Edge / Opera). Firefox / Safari fall back to CPU.
• —Mic input requires HTTPS — HuggingFace Spaces already serves over HTTPS.

Source

Full project, native C++ CLI, Python / Node bindings, and conversion scripts: github.com/RapidAI/RapidSpeech.cpp