README.md

metadata

title: kitsunechat
emoji: 🦊
colorFrom: blue
colorTo: purple
sdk: static
pinned: false
short_description: Voice-enabled Gemma 4 chat (Pocket TTS)
models:
  - litert-community/gemma-4-E2B-it-litert-lm
  - litert-community/gemma-4-E4B-it-litert-lm

kitsunechat

A voice-enabled fork of tylermullen/Gemma4 — same MediaPipe Gemma 4 chat UI, plus Pocket TTS running client-side. Every reply the model generates is automatically converted to speech and played back as it completes.

What's different

tylermullen/Gemma4 is a chat UI. This fork keeps everything the same and adds:

• tts-bridge.js — observes the chat history shadow DOM for completed model messages (the .redo-button only appears when doneGenerating=true, so it's the reliable "message is final" signal). When it appears, the message text is fed to Pocket TTS and the synthesized audio is attached inline as a <audio> element beneath the reply.
• worker.js + pkg/ + tts-client.js — the Pocket TTS engine (Rust→WASM). Same code as our tts-site, with the audio/audios segmenter bug fixed. Model weights + voice embeddings are fetched at runtime from HuggingFace, then cached in the browser's Cache API.
• Voice — azelma, by default.

The TTS runs entirely in your browser. Nothing leaves your device.

File layout

index.html          loads bundle.js + tts-bridge.js
bundle.js           MediaPipe Gemma 4 chat UI (from tylermullen/Gemma4)
tts-bridge.js       ← observes chat-history, queues completed messages for TTS
tts-client.js       orchestrator — picks worker per model
worker.js           Pocket TTS worker (loads /pkg/tts_wasm_bg.wasm)
audio-worklet.js    Web Audio worklet (not currently used; available for streaming)
pkg/                Pocket TTS WASM build (Rust → wasm-bindgen)
wasm/               MediaPipe LLM Inference WASM (from upstream)

Limits

• Pocket TTS text is bounded. Long LLM replies get split into segments; very long replies queue and play one after another.
• Browser autoplay policies may block the first auto-played clip until you click the page once. Subsequent replies autoplay normally.