Pocket TTS
Show HN: Pocket TTS running in (mobile) Safari
Last verified:
What is Pocket TTS?
Pocket TTS is a lightweight, open-source text-to-speech application developed by Kyutai Labs that generates high-quality audio from text entirely on your CPU without requiring GPUs or internet connectivity. Released in January 2026, this tiny model with only 100 million parameters is designed to fit in your pocket and run efficiently on any computer, making it accessible for developers, educators, accessibility advocates, and anyone needing offline speech synthesis.
The tool offers fully local, offline operation with no API calls, 8 built-in voices (alba, marius, javert, jean, fantine, cosette, eponine, azelma), voice cloning support from WAV samples, and low latency of approximately 200ms for the first audio chunk. It generates audio at 2-6x real-time speed on CPU, supports multiple languages including English, French, German, Portuguese, Italian, and Spanish, and provides both a simple Python API and command-line interface for easy integration into projects.
Pocket TTS is ideal for developers building screen readers, audiobook generators, Discord bots, Home Assistant integrations, Unity applications, and accessibility tools. It runs on Python 3.10-3.14 with PyTorch 2.5+ CPU version, uses only 2 CPU cores, handles infinitely long text inputs, supports audio streaming, and can even run client-side in browsers via WebAssembly implementations from the community.
Pocket TTS pricing
Pricing model: Freemium
Pocket TTS is completely free and open-source under the MIT license. There is no paid tier or subscription - the tool is available via pip install pocket-tts or uvx pocket-tts with no API costs since it runs entirely locally. The model is hosted on Hugging Face (gated, requires license acceptance) but downloading and using it is free. All features including voice cloning, all 8 built-in voices, multi-language support, and the Python API are included at no cost.
Pocket TTS pros
- Fully local and offline - no API calls or internet required
- CPU-only - no GPU needed, works on any computer
- Fast generation at 2-6x real-time on CPU
- 8 built-in voices with different characteristics
- Voice cloning from any WAV audio sample
- Low latency with ~200ms first audio chunk
- Simple Python API for easy project integration
- Small model size at only 100M parameters
- Uses only 2 CPU cores for efficiency
- Multi-language support including English, French, German, Portuguese, Italian, Spanish
- Handles infinitely long text inputs without limits
- Audio streaming capability for real-time output
- CLI and Python library options available
- Open-source with MIT license
- Can run in browser via WebAssembly community implementations
Pocket TTS cons
- Model is gated - requires accepting license on Hugging Face first
- English language only for base voices in v1
- First run downloads the model (~100M parameters)
- No support for adding silence/pauses in text input
- GPU execution shows no speedup over CPU
- Non-English languages have slower 24-layer high-quality variants
- Audio quality depends on quality of voice cloning sample
- Requires Python 3.10-3.14 and PyTorch 2.5+ installation