Questions and answers

What is Zero Talk 11M?

Zero Talk 11M is an English text-to-speech model with about 11M parameters (10.8M, plus about 94k for the voice). It reads English text aloud in one voice, and it runs on your own device: in a web browser through the model page and the npm package, or from the Python and Rust packages. The engine, the model and the text front end together are about 9 MB.

Why not the browser's built-in voices (speechSynthesis)?

The browser's built-in voices are free and need no files to load. They come from the operating system, so the voice you hear depends on the device and on the voices installed on it.

Zero Talk keeps its own voice in its model files, so it is the same voice on every device that runs the model. You can also save the audio: the page saves Ogg Opus, and the packages write WAV or Ogg Opus files. Zero Talk needs about 9 MB of files, which the page fetches from this site when you tap Load the voice.

How does it compare with Kokoro-82M, Pocket TTS, Kitten TTS and Piper?

Each figure is as published on 2026-11-05, in the project's own README or model card, or in its Hugging Face, npm or GitHub listing. File sizes are the model files named, except Zero Talk's figure, which also includes its engine and text front end. None of these figures is a quality claim. Each of the other projects offers more voices or languages than Zero Talk, which is one English voice.

Kokoro-82M (Apache-2.0, 82 million parameters) has 54 voices in 8 languages, according to its model card. Its weights file, kokoro-v1_0.pth, is 327,212,226 bytes, and each voice file is about 0.5 MB. Its official browser package is kokoro-js on npm, whose npm record names the hexgrad author and the hexgrad/kokoro repository. A community ONNX conversion, onnx-community/Kokoro-82M-ONNX, has eight ONNX model files of 86,033,068 to 325,525,037 bytes.

Pocket TTS (Kyutai) has its code under the MIT licence, and its Hugging Face model card gives CC-BY-4.0 for the model. Its README gives a main model of 100M parameters and English models of 18.2M, 6.9M and 1.9M parameters; the 18.2M and 6.9M English model files are 111,728,528 and 66,482,920 bytes. It has 26 pre-made voices, supports voice cloning, and covers seven languages: English, French, German, Portuguese, Italian, Spanish and Dutch. It names no official browser package; its README lists browser implementations by other authors.

Kitten TTS 0.8 has four ONNX models of 15M to 80M parameters, each under Apache-2.0 on Hugging Face; the code is Apache-2.0 too. The 15M int8 model's ONNX file is 24,369,971 bytes, with a 3,278,902-byte voices file, and the 80M mini model's ONNX file is 78,268,016 bytes. The 0.8 models have eight voices, and their documentation lists no languages. Its file list has no browser package. Kitten TTS 2, a separate 1.7B-parameter model with 115 voices and 20 languages, is under the Stellon Labs Community License.

Piper has 178 voices in 53 languages, as the voices.json index of rhasspy/piper-voices lists them. Its engine repository is under GPL-3.0; the earlier rhasspy/piper repository is archived and was under MIT. The voices are under MIT on Hugging Face, and each voice is a separate file: en_US-lessac-medium.onnx is 63,201,294 bytes. Its README names no browser package, and its file list has no browser or WebAssembly files.

Of the other four projects, only Kokoro-82M names an official browser package in the pages above (kokoro-js). Zero Talk's official browser package is the npm package @labzero-ai/zero-talk, which also runs in Node.js.

Why a custom file format, and can the model be loaded in PyTorch or ONNX?

The model is in model.bin, a file in the engine's own format, which the Zero Talk engine reads. It is not a PyTorch or ONNX file. To run the model, install the npm, Python or Rust package: each one includes the model and reads it for you.

How do numbers, dates and abbreviations sound?

Numbers, dates and prices written as digits are read out. Some names are read oddly. This page does not describe how abbreviations are read; the limits list on the model page is the reference for what does not work well.

Which browsers work?

Tested in Chromium on an Apple-silicon laptop and Safari 26 on an iPhone 17 Pro, with both engines. Other browsers and phones are untested. The page uses WebGPU where the browser offers it on a secure page, and WebAssembly where WebGPU is missing or fails.

Is the audio marked as AI-generated?

Yes. WAV and Ogg Opus files written by Zero Talk carry a marker in their metadata: a comment saying the speech is AI-generated and made by Zero Talk 11M and, in Ogg files, the IPTC digital source type for AI-generated media. The audio samples are not changed. Raw PCM output has no container and cannot carry a marker.

What about espeak-ng and its licence?

espeak-ng was among the tools used to prepare the front end's pronunciation data. It is not included, and the program does not run it. Its licence is the GNU General Public License, version 3 or later (https://github.com/espeak-ng/espeak-ng). See the notices.

What is Zero Talk built on, and what does the licence allow?

Zero Talk builds on Kokoro-82M (Apache-2.0), whose author is hexgrad. Its voice is Kokoro's af_heart. Zero Talk, its model files and its engine are licensed under the Apache License 2.0, which allows commercial use. If you redistribute the files, keep the licence text, the NOTICE and the third-party notices.

Does anything leave your device?

No. The page makes no third-party requests, and the text you type is not sent anywhere. The only requests the page makes go to this site: the model files come from here when you tap Load the voice.

Who is behind this?

Lab Zero is an independent lab. The model, the engine and the measurements are public; see the notices, the repository and model.json.