A hands-free AI voice assistant that runs entirely in your browser. Speech recognition, LLM, and text-to-speech all run locally using WebGPU - no API keys, no server, no data leaves your device. Just talk naturally and the AI responds.

Try it now: HuggingFace Space

Everything runs in your browser:

• Speech-to-Text: Whisper model via WebGPU/WASM
• Voice Activity Detection: Silero VAD detects when you're speaking
• LLM: Qwen 1.5B via WebLLM (easily swappable - see below)
• Text-to-Speech: Supertonic TTS with 10 natural voices

No audio leaves your device. No API keys needed. Just open and talk.

The built-in LLM is just a demo. The real value is the voice pipeline - STT, VAD, and TTS all wired up and working. Rip out the tiny in-browser model and point it at any LLM you want:

• Claude, GPT-4, Gemini - via API routes
• Ollama, LM Studio - local inference servers
• Any OpenAI-compatible endpoint

It's ~10 lines of code to swap. See Using a Different LLM below.

# Install dependencies
pnpm install

# Run development server
pnpm dev

Open http://localhost:3000 in Chrome or Edge.

First load downloads ~1GB of models from HuggingFace CDN. After that, everything runs offline.

• Browser: Chrome 113+ or Edge 113+ (WebGPU required)
• RAM: ~4GB available for models
• Microphone: Required for voice input

Falls back to WASM if WebGPU unavailable (slower but works everywhere).

┌─────────────────────────────────────────────────────────────┐
│                         Browser                             │
│                                                             │
│  Microphone                                                 │
│       |                                                     │
│       v                                                     │
│  ┌──────────┐   ┌──────────┐   ┌──────────┐   ┌──────────┐ │
│  │ Silero   │ > │ Whisper  │ > │ WebLLM   │ > │Supertonic│ │
│  │ VAD      │   │ STT      │   │ (Qwen)   │   │ TTS      │ │
│  └──────────┘   └──────────┘   └──────────┘   └──────────┘ │
│       |              |              |              |        │
│  Detects        Transcribes    Generates       Speaks      │
│  speech         to text        response        response    │
│                                                             │
└─────────────────────────────────────────────────────────────┘

src/
├── app/
│   ├── page.tsx              # Main voice chat UI
│   ├── layout.tsx            # App layout
│   └── globals.css           # Styles
├── components/ui/            # UI components
├── hooks/
│   ├── use-webllm.ts         # WebLLM integration
│   └── use-tts.ts            # TTS integration
└── lib/
    ├── tts.ts                # TTS pipeline
    └── splitter.ts           # Text chunking

public/
├── stt-worker-esm.js         # Whisper + VAD worker
├── vad-processor.js          # Audio worklet
└── voices/                   # TTS voice embeddings (bundled)

This demo uses WebLLM for fully local operation. To use an external LLM instead:

1. Create an API route (e.g., src/app/api/chat/route.ts)
2. In page.tsx, find handleLLMResponse() and replace the WebLLM call:

// Instead of webllm.chat(), call your API:
const response = await fetch("/api/chat", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({ messages: conversationHistory })
});
const data = await response.json();
return data.response;

• Framework: Next.js 16, React 19
• STT: Whisper via @huggingface/transformers
• VAD: Silero VAD via ONNX Runtime
• LLM: Qwen 1.5B via @mlc-ai/web-llm
• TTS: Supertonic via @huggingface/transformers
• Styling: Tailwind CSS v4

10 voices bundled (5 female, 5 male):

• F1: Calm, steady
• F2: Bright, cheerful
• F3: Professional
• F4: Confident
• F5: Gentle
• M1: Lively, upbeat
• M2: Deep, calm
• M3: Authoritative
• M4: Soft, friendly
• M5: Warm

MIT License - see LICENSE

• Whisper - OpenAI
• Silero VAD - Silero Team
• WebLLM - MLC AI
• Transformers.js - Hugging Face
• Supertonic TTS - Supertone