Pocket TTS

Pocket TTS is a text-to-speech application from Kyutai that runs on CPUs without needing a GPU or a hosted web API. Users generate audio from text through a Python library, a command-line interface, a local HTTP server, or a browser demo on the Kyutai website.

What It Is

Pocket TTS is a roughly 100M-parameter text-to-speech model with a Python package (pip install pocket-tts) and a CLI. The project lists audio streaming, voice cloning, and support for English, French, German, Portuguese, Italian, Spanish and Dutch. It says the first audio chunk arrives in about 200ms and that it runs about 6x faster than real time on a MacBook Air M4 CPU using two cores. It also states that it can handle arbitrarily long text input.

Workflow and Commands

The generate command writes a wav file from text, with options for voice, language and custom config. The serve command starts a local web interface and HTTP server that keeps the model in memory between requests. The export-voice command converts an audio sample into a safetensors voice embedding that loads quickly. A catalog of pre-made voices is included, and --voice also accepts a wav file for cloning. In Python, TTSModel.load_model() and generate_audio() return PCM audio as a tensor.

Deployment Options

Each release ships standalone binaries for Linux, macOS and Windows that need neither Python nor PyTorch. These are a Rust port that runs on CPU only. GPU use is not officially supported in the Python library, though the docs describe moving the model to CUDA manually and report a speedup on one tested T4 setup. Adding silence for pauses is listed as unsupported.

Community Ecosystem

The docs list community-trained models in languages such as Czech, Hindi, Korean, Japanese and Welsh, loadable through the --config option. They also list in-browser ports, alternative implementations and projects built on Pocket TTS. The docs include a prohibited-use policy covering voice cloning without consent and deceptive content.