| --- |
| title: WebASMR |
| colorFrom: indigo |
| colorTo: purple |
| sdk: static |
| pinned: false |
| --- |
| |
| # WebASMR |
|
|
| Bilingual (Chinese-first) ASMR generation, running entirely in the browser. |
|
|
| ## What works in this build |
|
|
| **TTS only.** Type Chinese (or English) text, pick a voice, hit 合成. The page |
| loads F5-TTS ONNX from `DownFlow/webasmr-f5tts`, runs the flow-matching loop and |
| the vocoder in a Web Worker, and plays the result. The stats block shows |
| wall-clock time and level, which is what we want measured on real hardware. |
|
|
| The DiT is the **4-step reflow-distilled, CFG-free** model (4 int4 Euler steps, |
| sway -1). First click downloads ~240 MB of weights (int4 DiT bundle + int4 |
| vocoder), cached afterwards. On a machine with hardware WebGPU the DiT runs on |
| hand-written WGSL kernels (about 1.2 s for a short line); otherwise it falls |
| back to ONNX Runtime on wasm (around half a minute). |
|
|
| ## What is not here yet |
|
|
| The **LLM director** (script generation, genres, radar map, voice/artist picker, |
| breath rings) needs the hybrid decode graph, which is still being exported. |
| Until then this Space measures the acoustic stack only. |
|
|
| Design notes: `docs/webml_app.md` in the project repo. |
|
|
