Pocket TTS
A small text-to-speech model that runs on a regular CPU, no GPU needed.
- Stars
- 10.2k
- Forks
- 1.1k
- Licence
- MIT
- Language
- Python
- Last update
- 6 days ago
- Type
- App
What it is
Pocket TTS turns written text into spoken audio. It is built to run on an ordinary CPU rather than needing a graphics card or a paid web API. You can use it from the command line, from a local HTTP server, or from Python code, and it can clone a voice from a short audio sample.
- Generate a wav file from text with the generate command
- Run a local server and send text to it over HTTP
- Clone a voice from a sample wav file using the --voice option
- Stream audio so the first chunk arrives in roughly 200ms according to the README
Great for
- Developers who want speech generation inside a Python project without a GPU
- People who want to generate narration or voiceovers locally on a laptop
- Anyone who wants to try voice cloning from their own wav or mp3 file
- Users who prefer offline or private generation over a cloud service
Not ideal for
- Users who need many languages beyond the listed seven, since the README says more may come later
- Projects that need the highest possible audio quality from a large GPU model, because the README presents this as a small, CPU-focused model
- People who want a no-install option only, though the README does link a website demo for that
How to get it
- 1
Install with a package manager
You'll need: Python 3.10 or newer · PyTorch 2.5+ (CPU build recommended on Linux) · uv (optional, recommended by the README)
Install with pip
pip install pocket-tts - 2
Run it
- Run `uvx pocket-tts generate` or `pocket-tts generate` after installing
- Find the output file at ./tts_output.wav
- Change the text with --text or the voice with --voice
- Run `pocket-tts serve` and open http://localhost:8000 for the web interface
- 3
Get the most out of it
- Use the serve command when trying many voices or prompts, since the model stays loaded in memory
- Use export-voice to convert a voice sample once, because loading the exported safetensors file is fast
- Try the --language option to switch between English, Italian, and other supported models
- Clean voice samples before cloning, as the README notes sample quality carries over into the output
Downloads
Strengths
- Runs on CPU and the README says it uses only 2 CPU cores
- Has a Python API, a CLI, and standalone binaries that need no Python
- MIT licensed
Trade-offs
- The README's performance figures come from the authors and a specific MacBook Air M4 setup, so results on your hardware may differ
- Voice cloning quality depends on the input sample, and the README recommends cleaning it first
- Supports seven languages, with more only mentioned as a future possibility
Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.