Loading tool…
Paste an article, script or lesson and hear it read aloud by a natural-sounding voice. Kokoro runs on your own device, so you can generate as much speech as you like and download it as MP3 or WAV, with no account or character limits.
Processing locally on your device
Overview
Kokoro is an 82-million-parameter text-to-speech model released under Apache-2.0. Despite its small size, it produces speech that compares well with much larger systems. Text is first normalised (numbers, currencies, abbreviations), then converted into phonemes with eSpeak NG, and the model turns those phonemes plus a voice style into a 24 kHz waveform. This tool uses the kokoro-js library for the text front end and voices, running the model on the same self-hosted ONNX Runtime as the other AI tools.
There are 28 English voices: 20 American and 8 British, female and male. Each has a quality grade from the model's authors; Heart and Bella are the most polished. You can preview any voice with a short sample before generating, and set the speaking speed from half to double. Kokoro speaks English only, so text in other languages will be read with English pronunciation.
Long text is split into sentences, and very long sentences are split at commas and semicolons, because the model reads at most about 500 phonemes at a time. Each piece is generated in turn with a progress indicator, short pauses are added between sentences and longer ones between paragraphs, and the finished audio plays with the current sentence highlighted. Click any sentence to jump to it.
On the CPU the 8-bit model is a 92 MB one-time download and needs roughly one to three seconds of work per second of speech, depending on your processor. With a WebGPU graphics card you can switch to the full-precision model (326 MB), which is usually faster. WAV downloads instantly; MP3, M4A, OGG, Opus and FLAC are encoded with FFmpeg in the browser.
| Voice | Accent | Gender | Grade |
|---|---|---|---|
| Heart | American | Female | A |
| Bella | American | Female | A- |
| Nicole | American | Female | B- |
| Emma | British | Female | B- |
| Michael / Fenrir / Puck | American | Male | C+ |
| George / Fable | British | Male | C |
Type or paste English text, up to 20,000 characters.
Choose a voice and click Preview to hear it.
Set the speed, and enable the graphics card if your browser supports WebGPU.
Click Generate speech and follow the sentence-by-sentence progress.
Play the result, then pick a format and download it.
Use cases
Create narration for explainer videos, product demos and social clips without recording your own voice.
Turn long articles, reports or study notes into audio you can listen to while commuting or exercising.
Produce consistent narration for course slides and lessons, and regenerate a sentence whenever the script changes.
Mock up voice prompts for apps, games and phone menus before booking a voice actor.
Pro tips
Answers to common questions about the text to speech.
The Kokoro model is licensed under Apache-2.0, which permits commercial use. Check that you have the rights to the text you're reading, too.
Not yet. The voices here are English only (American and British). Other languages will be pronounced as if they were English.
On a typical laptop CPU expect around one to three seconds of processing per second of speech. A WebGPU graphics card is usually faster. Progress is shown sentence by sentence.
Up to 20,000 characters per run, roughly 20–25 minutes of speech. For longer texts, generate in parts.
The model (92 MB on CPU) and the voice data download once. After that they're cached in your browser and load quickly, even offline.
No. ToolEnhancer processes files on your own device inside the browser. They are never sent to, stored on or seen by our servers.
Yes. The text to speech is free to use with no sign-up, no watermarks and no daily limits.