hexgrad: Kokoro 82M

hexgrad: Kokoro 82M is a lightweight, open-weight text-to-speech model from hexgrad. It sits in the speech category, so its job is narrow and clear: turn written text into spoken audio. Because the weights are open, you can reason about deployment and hosting more freely than with a closed hosted voice service, and the 82M parameter count signals a small model rather than a large general purpose one.

The model takes text as its only input and returns speech as its only output. It has a context window of 4096 tokens and a maximum output of 3686, which bounds how much text you can synthesize in a single request. Voice selection is broad: the model ships with a large set of named voices covering American English, British English, Spanish, French, Hindi, Italian, Japanese, Portuguese and Chinese, with both female and male options in most of those languages. Names such as af_heart, am_onyx, bf_emma, em_alex, jf_alpha and zf_xiaoxiao are examples from that list.

Those capabilities point to a fairly specific set of uses. Narration and voiceover for articles, documentation or short videos fit well, since you can pick a voice that matches the target language and audience. Accessibility features such as reading text aloud in an app, or generating audio versions of written content, are natural fits. Localization work benefits from the multi-language voice set, letting you produce the same script in several languages without switching providers. Conversational agents and interactive tools can use it for spoken replies, and because the model is small, it is a reasonable candidate for self-hosted or on-device style deployments where a large hosted TTS service would be overkill.

You call it through api14's OpenAI-compatible API using the model id "hexgrad/kokoro-82m". Because the interface follows the OpenAI format, existing clients and SDKs that already speak that protocol can point at api14 with minimal changes. The model reports no supported request parameters beyond the basics, so plan on sending your text and your chosen voice rather than relying on tuning knobs like temperature or top_p. Check the live price table on the page for current rates.

The main constraint to keep in mind is scope. This is a text-to-speech model, not a chat or reasoning model, so tasks like summarization, translation or question answering belong with a text model instead. The 4096 token context and 3686 token output ceiling also mean long documents need to be split into chunks before synthesis. If you need speech to text rather than text to speech, or a model that handles audio understanding, look elsewhere.