README.md

metadata

title: Pocket-TTS 100M
emoji: 🔊
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 6.2.0
app_file: app.py
pinned: true
license: apache-2.0
short_description: High quality, efficient voice cloning. Just 100M parameters.

Pocket-TTS

A lightweight text-to-speech application built with kyutai/pocket-tts and Gradio.

Features

• Fast CPU inference — ~6x faster than real-time on modern CPUs
• Low latency — ~200ms to first audio chunk
• Streaming output — Audio plays as it generates
• Voice cloning — Use custom voice samples (MP3, WAV, FLAC, etc.)
• Pre-computed embeddings — Voices work without voice cloning auth on HF Spaces

Quick Start

pip install -r requirements.txt
python app.py

Open http://127.0.0.1:7860 in your browser.

Adding Custom Voices

1. Drop audio files (MP3, WAV, etc.) into the voices/ directory
2. Restart the app
3. Embeddings are created automatically on first boot (requires HF auth locally)
4. Once created, embeddings are saved to embeddings/ and work without auth

Structure

Pocket-TTS/
├── app.py
├── requirements.txt
├── voices/           # Your custom voice audio files
│   └── my_voice.mp3
└── embeddings/       # Auto-generated (commit these for HF Spaces)
    └── my_voice.safetensors

HuggingFace Spaces Deployment

Option 1: Pre-commit embeddings (no auth needed on Space)

1. Run the app locally first (with HF auth) to generate embeddings
2. Commit both voices/ and embeddings/ directories
3. The Space will use pre-computed embeddings

Option 2: Auto-create embeddings on Space (requires valid token)

1. Accept terms at https://huggingface.co/kyutai/pocket-tts
2. Add HF_TOKEN secret in Space settings (must be a valid token)
3. Embeddings are created automatically on first boot

Model Info

• Model: kyutai/pocket-tts
• Parameters: 100M
• Language: English only
• Sample rate: 24kHz

License

See the kyutai/pocket-tts model card for licensing information.