Kokoro TTS
Kokoro is an 82-million-parameter open-weight text-to-speech model. This page reads your text aloud in one of 54 preset voices and returns the speech as an MP3.
Overview
Kokoro is an open-weight text-to-speech model with 82 million parameters, published by hexgrad. Its architecture comes from StyleTTS 2 and ISTFTNet, and it is decoder-only, with no diffusion step. hexgrad describes its quality as comparable to larger models.
This page runs Kokoro v1.0, released on 27 January 2025 and trained on a few hundred hours of audio. That release has 54 voices across eight languages. A run reads your text in one voice and returns a single MP3. Longer text is split into segments that are joined back into one track.
What you can build
- Narration for explainer videos and e-learning modules
- Spoken versions of articles and help pages
- Voice replies for assistants and chatbots
- Read-aloud features for accessibility
- Placeholder voice-overs while a script is still changing
Inputs
- Text: the words to speak. The default is a one-sentence quotation.
- Voice: one of the 54 Kokoro v1.0 presets. Heart is the default. Others include Bella, Nicole and Michael for American English, Emma and George for British English, Siwis for French, and Sara and Nicola for Italian.
- Language: American English (default), British English, Spanish, French, Hindi, Italian or Brazilian Portuguese. This setting decides how the text is converted to phonemes before it is spoken.
Outputs
One mono MP3 at 24 kHz. The two samples on this page are single quotations read aloud, about 5 and 6 seconds long.
Recommended settings
Kokoro's voice list places each voice under one language, so match the two settings, for example Heart or Bella with American English and Emma with British English. hexgrad's voice notes grade Heart (A) and Bella (A-) highest among the American English voices.
The same notes say most voices do best on utterances of 100 to 200 tokens. They can sound weak below roughly 10 to 20 tokens and can rush above about 400. Line breaks in your text start a new segment, so a long passage split into short paragraphs stays closer to that range.
Limitations
- hexgrad notes that support for languages other than English may be absent or thin, because of weaker phoneme conversion or too little training data. French has a single voice.
- The voice list includes Japanese and Mandarin Chinese voices, but the language setting has no option for either language.
- Voices are fixed presets. This page has no voice cloning and no control over speaking rate.
- Some names appear more than once in the voice list (Alpha, Alex, Dora, Santa) because they belong to different languages.
Safety & compliance
The weights are released under the Apache 2.0 license. hexgrad states that Kokoro was trained only on permissive or non-copyrighted audio: public domain recordings, audio under licenses such as Apache and MIT, CC BY datasets, and synthetic speech from closed TTS models.
Example prompts
Great starting points for kokoro_tts.
API quick start
Run kokoro_tts with a single API call.
{
"prompt": "And, when you want something, all the uni…",
"voice": "af_heart",
"langCode": "a"
}curl -X POST "https://api.wiro.ai/v1/Run/wiro/kokoro-tts" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "And, when you want something, all the uni…",
"voice": "af_heart",
"langCode": "a"
}
JSONOther models you might like
Explore similar audio & speech models.



