Text to Speech Beta
Turn text into speech with a local AI voice model. No per-character fees.
How to text to speech
- Paste or type your text. The model downloads once on first use, roughly 48 MB.
- Pick a voice and adjust speed if needed.
- Generate and download the audio.
About this tool
Every commercial text-to-speech API meters you by the character, because the compute to synthesise speech is real and costs money on a server. Kokoro is an open-weights neural TTS model that runs well on consumer hardware. In a browser, it runs through WebAssembly and WebGPU, which means the speech synthesis happens on your own machine and the per-character cost is zero — because the cost is borne by your CPU.
The voice quality is meaningfully better than older open-source TTS systems. Kokoro uses a diffusion-based architecture that produces more natural prosody than the recurrent models that dominated the field a few years ago. English is strong. The model also supports French, Spanish, Japanese, Korean, Chinese, Hindi and a handful of other languages — but those voice models were trained on less data, and the quality gap versus English is noticeable.
A 48 MB model download is genuinely small for a capable neural voice model. It downloads once and is cached; subsequent uses start in under a second. The practical limit is your available RAM: very large batches of audio generation can accumulate in memory before the download is ready.