GUIDE

A Kokoro TTS API: self-hosted, third-party or a hosted alternative

Kokoro-82M is a popular open text to speech model: 82 million parameters, Apache-licensed weights and, in version 1.0, 54 voices in 8 languages. Its model card names no official hosted API from its author, so a “Kokoro API” means either running the model on your own server or paying a third-party host.

This guide compares those options with the SuperClonica speech API, which runs a different open model, Supertonic 3, as a hosted service with 190 catalog voices.

Get a free API key Hear the voices Or run the engine yourself

Option 1: host Kokoro yourself

The weights are Apache-licensed, so you can wrap the model in your own HTTP server and run it on a CPU or GPU. You keep full control and pay only for the machine, but you also maintain the server, scaling and updates. The model card warns that some websites using the Kokoro name are not owned by or affiliated with the model’s author.

Option 2: a third-party Kokoro host

The model card says Kokoro has been deployed in many commercial APIs and, as of April 2025, quoted market rates under $1 per million characters, citing about $0.65 and $0.80 per million characters at two hosts; that is roughly $0.04–0.05 per hour of audio. Check each host’s current price, limits and terms yourself.

Option 3: the SuperClonica speech API

SuperClonica runs Supertonic 3, an open model of the same size class, as a hosted API. It streams audio from about 300 ms in our measurements and gives access to 190 catalog voices made by authors (characters, narrators, announcers) plus voices you train from your own recording. Pricing is by time: a free key gives a minute a day, and Starter costs $5 a month for up to 216 hours, about $0.023 an hour when used to the limit.

Which one to choose

Choose Kokoro if you need Chinese, which Supertonic 3 does not speak yet, or Kokoro’s particular voices, and you are happy to host it or pick a host. Choose SuperClonica if you want many ready-made character and narrator voices, custom trained voices and a hosted streaming API without a server to maintain. Both models are small and run on a CPU, so you can also try each locally first.