Three speech demos in your browser, no server side: offline ASR, online ASR (mic + VAD), and offline TTS — all running locally via WebAssembly + WebGPU. The page loads a single shared WASM module and uses three independent contexts (one per tab) so you can hot-switch between ASR and TTS without rebuilding.
wasm-examples/
├── index.html # Tabbed shell (ASR Offline / ASR Online / TTS)
├── styles.css # Shared styles
├── app.js # Tab router
├── rapidspeech-bridge.js # JS wrapper around the WASM exports
├── rapidspeech-wasm.{js,wasm} # Build output (Emscripten)
├── utils.js # WAV decode / encode / format helpers
├── pages/
│ ├── asr-offline.js # Upload WAV or record → optional VAD → ASR
│ ├── asr-online.js # Mic capture → neural VAD (or energy gate) → ASR
│ └── tts-offline.js # text → synth → playback / download
└── serve.py # Tiny dev server with COOP/COEP headers
# 1. Build the WASM module (once; ~minutes the first time)
cd rapidspeech/wasm
./build-wasm.sh
# Produces rapidspeech-wasm.{js,wasm} → copied into wasm-examples/
# 2. Serve the demo (COOP/COEP required for pthreads + SharedArrayBuffer)
cd wasm-examples
python3 serve.py 8000
# 3. Open http://localhost:8000 — defaults to the ASR Offline tab.
# Direct-link a tab via #asr-online or #tts-offline in the URL.
⚠️ Useserve.py, notpython -m http.server. The plain HTTP server does not emit theCross-Origin-Opener-Policy/Cross-Origin-Embedder-Policyheaders that pthreads + SharedArrayBuffer require — without them the WASM module aborts at startup.
coi-serviceworker.js is loaded as the very first script in index.html. On
hosts that won't let you set COOP/COEP server-side (ModelScope Static Spaces,
GitHub Pages, some HuggingFace embeds), it registers a service worker that
intercepts every response and rewrites the headers itself. First visit reloads
the page once; subsequent visits start cross-origin-isolated. HTTPS is
required — service workers don't register on plain HTTP.
If your host does set COOP/COEP (the bundled serve.py, your own nginx,
etc.), the shim is a no-op — it detects crossOriginIsolated is already
true and skips registration.
- Paste a model URL (e.g. an HF GGUF) → Load Model.
- Optionally load a VAD GGUF (silero-vad / firered-vad) — when present, long
clips are pre-segmented by speech activity and each region is transcribed
separately with
[mm:ss → mm:ss]timestamps. - Source: upload a WAV or record a clip with the mic. Re-decode with LLM on/off using the Re-decode (LLM) button.
- Same model loader; add a neural VAD GGUF for best segmentation, or leave it blank to fall back to an RMS-energy gate.
- Two-pass checkbox: emits a CTC-greedy partial (grey, italic) the moment the segment closes, then updates the same line in place with the LLM-rescored result.
- A warning banner appears when the rolling-average RTF crosses 1.0 — your device cannot keep up with audio at the current settings. Hints suggest turning off LLM or two-pass.
- Load an OmniVoice or OpenVoice2 GGUF.
- Adjust
instruct, language, seed, and diffusion steps. Generated audio appears as an in-page<audio>element with a one-click WAV download. - Voice cloning: upload a reference WAV + its transcript (OmniVoice).
- Emscripten SDK (emsdk activated in the shell):
source /path/to/emsdk/emsdk_env.sh - The build script enables both WebGPU and pthreads:
WebGPU is only effective on Chromium-based browsers as of 2026. On Firefox / Safari the runtime silently falls back to CPU.
cmake -DRS_WASM_WEBGPU=ON -DRS_WASM_PTHREADS=ON ...
- Chromium 113+ for WebGPU (Chrome / Edge / Opera). WebGPU on Firefox is still
behind
dom.webgpu.enabled. Safari has it shipping but with quirks. - Cross-origin isolation (COOP+COEP) — see the
serve.pynote above. - Mic input requires HTTPS or
localhost.
The whole wasm-examples/ directory is a static site. Drop it on any host
that lets you set the COOP/COEP headers (HuggingFace Spaces "Static",
Cloudflare Pages with _headers, ModelScope Studios, S3+CloudFront with a
Lambda@Edge, etc.). Without COOP/COEP the module loads but pthreads aborts.
git clone https://huggingface.co/spaces/YOUR_USER/YOUR_SPACE
cd YOUR_SPACE
cp /path/to/RapidSpeech.cpp/wasm-examples/* .
git lfs install && git lfs track "*.gguf" && git add .gitattributes
cp /path/to/model.gguf .
git add . && git commit -m "RapidSpeech WASM demo" && git pushHuggingFace Spaces emit COOP/COEP automatically for static spaces.
The RapidSpeechWASM and RapidSpeechVAD classes in
rapidspeech-bridge.js are the only thing pages talk
to. They wrap the following C exports (defined in
rapidspeech/wasm/rapidspeech_wasm.cpp):
| Export | Description |
|---|---|
rs_wasm_init_ex(path, task_type, threads) |
Load model with explicit task |
rs_wasm_push_audio(ptr, n) |
Push float32 PCM (ASR / TTS reference) |
rs_wasm_push_text(text) |
Push UTF-8 text for TTS |
rs_wasm_push_reference_audio / text(...) |
Voice cloning |
rs_wasm_process() async |
Run one inference step (await it!) |
rs_wasm_redecode() async |
Re-run decoder only (2-pass ASR) |
rs_wasm_get_text / get_audio_ptr / get_audio_len |
Read outputs |
rs_wasm_set_use_llm / set_ctc_precheck / set_tts_params / set_tts_diffusion_steps |
Knobs |
rs_wasm_vad_init / push_audio / drain_segments / drain_frames |
VAD streaming API |
process / redecode are bound with cwrap({async: true}) because they
suspend through WebGPU's queue ops under ASYNCIFY — always await them or
you'll get a Promise instead of a result.
The same rapidspeech-bridge.js runs under Node.js. See
../node-api-example/ for a CLI that loads GGUF
files from disk and runs ASR / TTS without a browser.
| Symptom | Likely cause |
|---|---|
GET /rapidspeech-wasm.wasm net::ERR_CONNECTION_REFUSED |
serve.py not running. |
SharedArrayBuffer is not defined / pthreads abort |
Missing COOP/COEP headers — don't use python -m http.server. |
Queue work failed with status 3 |
(Old bug) VAD on WebGPU. Already fixed — VAD is pinned to CPU in rs_wasm_vad_init. |
gguf_init_from_file_ptr: tensor name … is too long: 68 >= 64 |
WASM build was configured without GGML_MAX_NAME=128. Rebuild with rapidspeech/wasm — its CMakeLists adds this define. |
| RTF warning banner | Your device can't keep up at the current settings. Turn off LLM, turn off two-pass, or shorten the VAD silence window. |
../node-api-example/README.md— same WASM module, Node CLI../python-api-examples/README.md— native Python bindings (faster than WASM)../README.md— project overview and C++ CLI usage