Text-to-speech · v1
Zero Talk 11M
An English text-to-speech model with about 11M parameters. It runs in your browser, on this device, and nothing you type leaves it.
Read out loud:
Model details
- parameters
- 11M
- download
- 9 MB
- real time, iPhone 17 Pro
- 2.5×
- audio, mono
- 24 kHz
What it is
Zero Talk reads English text aloud in one voice. It builds on Kokoro-82M (see Credits), reworked into a model of about 11M parameters that streams. Its engine is written in Rust and compiled to WebAssembly for this page. Where the browser offers WebGPU, the neural networks run on the graphics processor and the audio starts playing while the rest is still being made. Elsewhere, WebAssembly does all the work and plays each sentence when it is done.
Specifications
- Parameters
- about 11M (10.8M, plus about 94k for the voice)
- Voice
- English, one voice (Kokoro's af_heart)
- Audio
- 24 kHz, mono; the demo saves Ogg Opus
- Download
- about 9 MB, fetched only when you tap
- Runs on
- this device, in a background thread: WebGPU where the browser offers it on a secure page, otherwise WebAssembly. No GPU needed.
- Tested in
- Chromium on an Apple-silicon laptop and Safari 26 on an iPhone 17 Pro, with both engines. Other browsers and phones are not tested yet.
- Version
- v1
Machine-readable facts: model.json.
Faster than real time on an iPhone 17 Pro
On an iPhone 17 Pro, an earlier build of this page made 8.0 s of speech in about 3.2 s, about 2.5 times faster than real time (one run, 2026-10-29). After each sentence, the line under the demo shows the same figure for your device.
Intended use and limits
Made for reading short English texts aloud where size and privacy matter. Not for long narration, other languages, or imitating real people. One voice, English only. Questions come out flat, and exclamations do not come across. Some S sounds have a slight lisp. Long sentences can lose their rhythm, and sentences that open with a greeting and a comma ("Hi, …") can sound unnatural, so short, plain sentences work best. We judged this by listening, in a small number of trials, not with many listeners.
All known limitations
- Very short utterances ("Hi.") can sound slightly cut off at the start.
- No controls for tempo or pauses.
- Accented Latin words are read in plain spelling; other scripts are left out with a warning.
- Some names are read oddly.
- The two browser engines give close but not identical audio.
What it does not do
Some of these are planned, some are reserved for a paid version, and some are not possible with this model.
Reserved for an upcoming paid version
- Control over how it speaks: expressive delivery styles and emphasis on chosen words.
- A second voice.
Not shipped yet
- The command-line program: built and measured, not offered yet.
- Natural intonation for questions: a known gap we plan to retrain.
- Confirmed support in Firefox, Safari on macOS, Android and mid-range phones: not tested yet.
Not possible with this model
- Other languages: it reads English only.
- Voice cloning: the voice is fixed and cannot be copied from a recording.
Credits
Zero Talk builds on Kokoro-82M (Apache-2.0), whose author is hexgrad. Its voice is Kokoro's af_heart. Its sound generator started as Kokoro's own, was compressed to about a quarter of its parameters, then retrained to stream. The network that turns text into sound learned from Kokoro as its teacher.
The streaming design, the 4-bit compression, the pronunciation network and the Rust engine are Lab Zero's work. The text rules are ported from misaki (Apache-2.0). The text front end also uses data from spaCy and en_core_web_sm (MIT), LibriSpeech transcripts (CC BY 4.0) and Unicode 15.0.0.
Third-party notices
Attributions and licence texts for the third-party material that Zero Talk uses: THIRD-PARTY-NOTICES.txt, with the NOTICE and the licence text.
Version history
v1: first version, one English voice.