README.md

---
title: PrimeTTS Streaming
emoji: 🔊
colorFrom: green
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
short_description: Real-time streaming zh-TW + English TTS (torch-free)
models:
  - Luigi/PrimeTTS
  - unsloth/gemma-3-270m-it-GGUF
  - 52AI/tinyllamas_zh
---

# PrimeTTS · real-time streaming TTS

Type text, hear audio **stream chunk-by-chunk** (24 frames ≈ 384 ms each) — first
sound in ~100 ms instead of waiting for the whole utterance.

- **Model**: token-level MB-iSTFT-VITS (band-attention encoder + causal vocoder),
  split into `enc` (text→z, once) + `dec` (z-chunk→audio, per chunk, overlap-save).
- **Backend**: [sherpa-onnx fork](https://github.com/vieenrose/sherpa-onnx/tree/mbistft-streaming-tts)
  `OfflineTtsMbistftStreamModel` (C++/ONNX Runtime), bit-exact to the reference.
- **Frontend**: g2pw (bopomofo, Taiwan readings) + g2p_en — **torch-free** (ONNX Runtime only).
- **UI**: FastAPI + WebSocket + Web Audio API (PCM chunks played as they arrive).

zh-TW + English, 16 kHz.