Kokoro
A small open-weight text-to-speech model with Apache-licensed weights that runs on modest hardware.
Instead of ElevenLabs
Screenshot from huggingface.co/hexgrad/Kokoro-82M, taken 2026-10-30, shown to help you recognise the tool. It belongs to its maker. Ask us to remove it.
Bigger usually sounds better in speech models, and Kokoro makes a pointed argument against that with 82 million parameters. Its model card says the weights are Apache-licensed, that you can deploy them in real use cases, and that version 1.0 from January 2025 handles 8 languages and 54 voices. You run it with a few lines of Python after installing the kokoro package and espeak-ng.
Kokoro scores a 5 on footprint and lock-in, a 4 on self-direction, power and delivery and a 3 on longevity.
The model card credits the StyleTTS 2 architecture from Li and colleagues and says it was trained by a person who goes by @rzvzn on Discord, so a very small group holds this together and Longevity is a 3. It warns that sites with kokoro in their domain are not affiliated with it. A pip install and a bit of Python is the entry price, and we have not listened to enough of it to rate voices, so Delivery rests on the project's own claims.
- An 82 million parameter model with Apache-licensed weights, 8 languages and 54 voices
- Trained by one person using the StyleTTS 2 architecture
- Needs Python and espeak-ng to run
What stands out
Do you make Kokoro? Get the badge for your site
This badge reports a dated review. It is not a certification and it costs nothing. Link it to this page and leave the numbers alone: the image always shows the current score.
For dark backgrounds, use kokoro-dark.svg in place of kokoro-light.svg.
For comparison, we also reviewed ElevenLabs (16/30) as a benchmark.
Other tools for “hear text read aloud without a subscription”
eSpeak NGA compact speech synthesizer that covers more than a hundred languages and accents.
PiperA fast, local neural text-to-speech engine you install with one command.
SherpaTTSAn Android text-to-speech engine that uses Piper or Coqui voices and reads aloud with no connection.
ChatterboxAn MIT-licensed text-to-speech model with voice cloning, from a company that sells voice AI.