Kokoro is an open-weight text-to-speech model with 82 million parameters. Despite a lightweight architecture based on StyleTTS 2 and ISTFTNet, it reaches quality comparable to larger models while running much faster and at lower cost. The weights are Apache-licensed, so the model can be deployed from personal projects to production systems. Version 1.0 was published in January 2025 with 54 voices across 8 languages, and the model is free to download from Hugging Face. Third-party APIs serve it for under $1 per million characters of input.
Kokoro is built for developers, researchers and product teams that want natural-sounding speech without paying for a large proprietary service or a heavy GPU. Indie developers can run it locally in Python after a short pip install, while companies can self-host it for predictable costs and data control. Because it was trained exclusively on permissive and non-copyrighted audio, teams with licensing concerns get a clearer provenance story. Hobbyists can try the hosted demo space first, then move to the open model in their own applications once they are happy with the voices.
Typical use cases include generating audiobook-style narration, adding voice output to chatbots and assistants, producing voice-overs for videos and e-learning, and reading articles aloud in apps. A developer can call the Python pipeline with a language code and a voice such as af_heart, then write the audio to a WAV file at 24,000 Hz. Other teams use hosted providers when they prefer an API over self-hosting. One minute of speech typically needs about 1,000 characters of input text. The project warns that websites using kokoro in their domain name are not affiliated with the model.
