Pocket TTS

Pocket TTS

A 100M-parameter text-to-speech model from Kyutai that runs on CPUs in six languages, streams audio with about 200 ms to the first chunk, and works from Python or the command line.

Made by kyutai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Pocket TTS is a lightweight text-to-speech application from Kyutai that runs on ordinary CPUs. You install it with pip and generate audio with a function call or one command, without a GPU or a speech web service.

It has about 100M parameters, streams audio, and handles text of any length. The card says it starts in roughly 200 ms and runs about 6 times faster than real time on a MacBook Air M4 CPU, using only 2 cores. This repository's name says it leaves out voice cloning.

Who it suits: Developers and tinkerers who want local speech output with no GPU, in English, French, German, Portuguese, Italian or Spanish. Check each voice's license page before using it commercially.

Industry | Who uses it | How they use it | Example
Software | App developers | Add local text-to-speech to an app | A reader that speaks without a server
Accessibility | Tool builders | Read text aloud on ordinary laptops | A screen-side reader on a CPU-only machine
Media | Content creators | Narrate short scripts | A quick voice-over draft
Education | Language-app makers | Speak example sentences in six languages | Pronunciation clips for learners

You can skip it if you need studio-grade expressive voices, or other languages, since the card lists six.

Where to use it: It runs on CPUs, from Python or a command line, and the card says it can also run in a browser.

• Python 3.10 to 3.14 with PyTorch 2.5 or newer
• The pocket-tts command line
• A browser, through the client-side versions the card lists
• A CPU-only PyTorch build is enough

Use case | Who | What they get back
Local narration | Developers | A wav file from text, with no GPU
Streaming speech | App builders | Audio chunks starting in about 200 ms
Multilingual voice output | Product teams | Speech in six languages

How to set it up: You need Python and PyTorch 2.5 or newer, but not the GPU build of PyTorch. The card recommends uv, which installs dependencies on the fly, or you can install the package with pip.

• Python 3.10 to 3.14
• PyTorch 2.5 or newer, CPU build is fine
• uv, or pip

1. Install uv, or plan to install the package with pip.
2. Run the generate command, or install the package first.
3. Pick a voice, text and language with the command options.

pip install pocket-tts

pip install -U huggingface_hub
hf download kyutai/pocket-tts-without-voice-cloning

How to use it: Run the generate command. It writes tts_output.wav with the default text and voice, and prints some speed statistics. Add options to change the voice, text and language.

uvx pocket-tts generate
# or if you installed it manually with pip:
pocket-tts generate

Tool | What it does
pocket-tts generate | Writes a wav file from text and shows speed statistics
--voice | Chooses the voice
--text | Sets the text to speak
--language | Picks a language model, for example italian_24l for a larger Italian one
--config | Points to custom weights, as a file, web or hf path

Example: You run the generate command with a voice such as alba and a short English sentence. The command saves tts_output.wav and prints speed statistics so you can see how fast it ran.

For newcomers and experts: If you are new, run the command with no options first, listen to the file, then change the voice. The card lists a small catalog, for example alba for English and estelle for French.

If you know the field, the card offers larger 24-layer variants for non-English languages that sound better but run slower. A training code release lets you train your own models, and a community section lists models others trained.

Install with uv or pip
Run pocket-tts generate
Text goes to the CPU model
Audio streams out in chunks
A wav file is written

Good to know: Each voice has its own license on a separate page, so check it. The repository name says it has no voice cloning, while the card's feature list mentions it, so confirm what this version supports before you rely on it.

• License: CC BY 4.0
• Cost to run: none named, since it runs on a CPU
• Data it reaches: the text you speak, plus the model and voice files you download
• Not for: other languages beyond the six, or voices without checking their licenses

On Hugging Face, 13 November 2026

• Library: pocket-tts
• Formats: Safetensors
• Downloads in the last 30 days: 12,179
• Likes: 44
• License: cc-by-4.0
• Languages: en, fr, de, pt, it, es
• Quantized versions: 16
• Fine-tunes: 4
• Tags: pocket-tts, safetensors

Links

• Model: https://huggingface.co/kyutai/pocket-tts-without-voice-cloning
• Files: https://huggingface.co/kyutai/pocket-tts-without-voice-cloning/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download kyutai/pocket-tts-without-voice-cloning

Then follow the usage in its model card. Nothing in this kit runs by itself.