---
title: PrimeTTS vs Inflect-Nano WASM
emoji: 🗣️
colorFrom: indigo
colorTo: pink
sdk: static
pinned: false
license: apache-2.0
short_description: 7-model on-device TTS A/B, 100% client-side (WASM)
models:
- owensong/Inflect-Nano-v1
- Luigi/PrimeTTS
- MediaTek-Research/Breeze2-VITS-onnx
- Luigi/vits-melo-tts-zh_en-8k
- Luigi/matcha-zh-tw-en-8k
---

# PrimeTTS vs Inflect-Nano — WASM edition

Fully client-side port of the original Gradio A/B demo. Pick a model, type text,
hit Generate — everything runs in your browser, no server.

| Model | Runtime |
|---|---|
| PrimeTTS v2.1 (3-voice), V2 Lite | onnxruntime-web, single-pass MB-iSTFT-VITS |
| PrimeTTS v1, Inflect-Nano-v1 | onnxruntime-web, FastSpeech + Snake-HiFiGAN (JS port of `host_regulate`) |
| Breeze2-VITS, MeloTTS, Matcha | [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx)'s official WASM TTS build (own compile per model — weights baked in at build time) |

**MOSS-TTS-Nano was removed from this demo.** It was ported (AR loop, RVQ
codec, from-scratch SentencePiece tokenizer, voice cloning from a built-in
preset or an uploaded reference clip) and worked, but as a 100M-parameter
autoregressive model it's a poor fit for a lightweight in-browser demo next
to the much faster VITS/FastSpeech options.

Model weights are fetched lazily from their original repos and cached by the
browser — nothing is bundled beyond the small text-processing assets.

## Local development

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