Kokoro is the open text-to-speech model people keep coming back to: only 82 million parameters, yet it sounds close to models ten times its size, and its Apache-2.0 license allows commercial use. The usual obstacle is the setup (Python, PyTorch, a phonemizer, eSpeak). On LiveWeights you can try it straight from the browser, with no install and no GPU to wait for.
Kokoro 82Mhexgrad · Apache-2.0One of the most popular open text-to-speech models: 82M parameters, natural voices in English, Italian, Spanish, French, Portuguese and Hindi, on a plain CPU.Try it online →What Kokoro-82M is
- 82 million parameters, based on the StyleTTS 2 architecture, released by hexgrad under Apache-2.0.
- 24 kHz audio, with 54 voices in the official v1.0 release.
- Languages: American and British English, Spanish, French, Italian, Brazilian Portuguese, Hindi, Japanese and Mandarin.
- Small enough for a CPU: we run the official 8-bit ONNX export (92 MB) on our server's processor, so there is no GPU cold start.
Voices you can try on LiveWeights
| Language | Voices |
|---|---|
| American English | Heart (the best rated), Bella, Nicole, Michael, Fenrir, Puck |
| British English | Emma, George |
| Italian | Sara, Nicola |
| Spanish | Dora, Alex |
| French | Siwis |
| Portuguese (Brazil) | Dora, Alex |
| Hindi | Alpha, Omega |
Each voice speaks one language: write your text in the language of the voice you pick. English goes through Kokoro's own phonemizer (misaki), which handles numbers, prices and abbreviations; the other languages use eSpeak, exactly like the official pipeline.
How to try it online
- Open the Kokoro 82M page and sign in, then pick a small credit pack (no subscription).
- Write up to 1,000 characters, pick a voice and a speed.
- Press Generate: the audio is ready in a few seconds, and you can listen, download the WAV or try another voice.
Kokoro vs other open TTS models
Kokoro has no voice cloning: if you need your own voice, or a voice designed from a description, try Qwen3-TTS or Chatterbox Multilingual. For tiny English models there are Phonon 31M and Paradee, which was distilled from Kokoro itself. We compared all of them in our text-to-speech roundup.
Then run it locally
The weights are on Hugging Face (hexgrad/Kokoro-82M). The official Python package is kokoro (it needs eSpeak NG for non-English languages); for a lighter setup without PyTorch there are ONNX exports (onnx-community/Kokoro-82M-v1.0-ONNX) that also run in the browser with Kokoro.js.