---
title: PrimeTTS Streaming WASM
emoji: 🔊
colorFrom: green
colorTo: indigo
sdk: static
pinned: false
short_description: Real-time streaming zh-TW + English TTS (WASM, no server)
models:
  - Luigi/PrimeTTS
---

# PrimeTTS · real-time streaming (WASM edition)

Fully client-side port of [Luigi/PrimeTTS-Streaming](https://huggingface.co/spaces/Luigi/PrimeTTS-Streaming)
(FastAPI + WebSocket) to a static Space that runs entirely in the browser —
no server, no WebSocket. Type text, hear audio **stream chunk-by-chunk**
(24 frames ≈ 384 ms each).

- **Vocoder**: token-level MB-iSTFT-VITS split into `enc` (text→z, once) +
  `dec` (z-chunk→audio, per chunk, overlap-save), run via
  [onnxruntime-web](https://github.com/microsoft/onnxruntime) (WASM).
- **Frontend**: g2pw (BERT ONNX, bopomofo/Taiwan readings) + CMU-dict-based
  English g2p — ported to plain JS, BERT inference via onnxruntime-web.
- **Streaming-input demo**: two GGUF LLMs (llama2.c-zh 15M story generator,
  Gemma-3-270m-it chat) run via [wllama](https://github.com/ngxson/wllama)
  (llama.cpp compiled to WASM); OpenCC (opencc-js) converts to zh-TW.
- **UI**: static HTML + Web Audio API, PCM chunks played as they arrive.

All model weights (~600 MB total: 2 ONNX vocoder splits, 1 ONNX g2p model,
2 GGUF LLMs) are fetched lazily from the
[Luigi/PrimeTTS](https://huggingface.co/Luigi/PrimeTTS) model repo on first
use and cached by the browser — nothing is bundled into this Space.

zh-TW + English, 16 kHz.

## Local development

```bash
npm install
npm run dev      # http://localhost:5173
npm run build    # outputs dist/
```