PocketTTS is a lightweight CPU text-to-speech launcher for Kyutai's PocketTTS model. It provides a Gradio Web UI for preset voices and optional voice cloning from a short audio sample.
- CPU-friendly PocketTTS speech generation.
- Eight built-in preset voices: Alba, Marius, Javert, Jean, Fantine, Cosette, Eponine, and Azelma.
- Optional custom voice cloning with a Hugging Face account that has accepted the
kyutai/pocket-ttsmodel terms. - Pinokio install, start, update, reset, disk deduplication, and Hugging Face login actions.
Intel Macs are not supported: pocket-tts requires torch 2.5 or newer, and PyTorch stopped publishing Intel macOS wheels after 2.2.2.
- Add this repository to Pinokio.
- Click
Install. - Click
Start. - Open
Open Web UIwhen Pinokio detects the Gradio URL.
Preset voices can run without Hugging Face login. Voice cloning requires model access:
- Accept the terms at https://huggingface.co/kyutai/pocket-tts.
- Click
Hugging Face Loginin the launcher. - Paste a Hugging Face token when
uvx hf auth loginprompts for it. - Restart PocketTTS if it was already running.
You can also paste a token in the Web UI's Hugging Face Token for Voice Cloning accordion.
install.js: installs Python dependencies fromapp/requirements.txt, then runstorch.js.start.js: launchesapp/app.pyin theapp/envvirtual environment and captures the local Gradio URL.hf-login.js: opens an interactive Hugging Face CLI login terminal for gated voice-cloning access.update.js: pulls the latest launcher code and upgrades app dependencies.reset.js: removes theapp/envvirtual environment, leaving the app source in place.link.js: deduplicates redundantapp/envfiles to save disk space.pinokio.js: renders the Pinokio menu dynamically based on install and running state.
PocketTTS/
├── app/
│ ├── app.py
│ └── requirements.txt
├── install.js
├── start.js
├── hf-login.js
├── update.js
├── reset.js
├── link.js
├── torch.js
├── pinokio.js
├── pinokio.json
├── .gitignore
├── README.md
└── icon.png
The launcher creates its virtual environment at app/env, which is gitignored.
Use app/env so a manual run shares the virtual environment the launcher creates.
Install torch before the requirements: pocket-tts depends on torch, so installing the requirements first pulls PyPI's default build (CUDA on Linux) and the CPU install afterwards is skipped as already satisfied.
On Windows:
python -m venv app\env
app\env\Scripts\activate
uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
uv pip install -r app\requirements.txt
cd app
python app.pyOn macOS or Linux:
python -m venv app/env
source app/env/bin/activate
uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
uv pip install -r app/requirements.txt
cd app
python app.pyA manual run listens on port 7860 unless you set PORT.
When the Gradio app is running, the generation function is exposed as the generate API route. The port is assigned dynamically by Pinokio — check the Open Web UI URL in the launcher for the actual port (e.g. http://127.0.0.1:PORT).
from gradio_client import Client
client = Client("http://127.0.0.1:PORT") # replace PORT with the port shown in the launcher
result = client.predict(
text="Hello from PocketTTS.",
preset_voice="Alba",
custom_voice_file=None,
api_name="/generate",
)
print(result)import { Client } from "@gradio/client";
const client = await Client.connect("http://127.0.0.1:PORT"); // replace PORT with the port shown in the launcher
const result = await client.predict("/generate", {
text: "Hello from PocketTTS.",
preset_voice: "Alba",
custom_voice_file: null
});
console.log(result);Gradio's HTTP API is two-step: the POST queues the job and returns an
event_id, then a GET on that id streams the result.
EVENT_ID=$(curl -s -X POST http://127.0.0.1:PORT/gradio_api/call/generate \
-H "Content-Type: application/json" \
-d '{"data":["Hello from PocketTTS.","Alba",null]}' \
| grep -o '"event_id":"[^"]*"' | cut -d'"' -f4)
curl -N http://127.0.0.1:PORT/gradio_api/call/generate/$EVENT_IDUse voice cloning only with explicit and lawful consent. Prohibited uses include impersonation without consent, misinformation, deception, and unlawful or privacy-invasive content.
- PocketTTS model: https://huggingface.co/kyutai/pocket-tts
- PocketTTS source: https://github.com/kyutai-labs/pocket-tts
- Research: https://kyutai.org/research/pocket-tts