DevToolPal
EN

Text to Speech

Turn English text into natural speech with the Kokoro-82M model running in your browser, then download it as WAV.

Runs locally, never uploaded

Speech model

Kokoro-82M runs entirely in your browser. The model downloads once from Hugging Face and is cached, so the next visit starts in seconds.

Options

Voices speak English only (American and British accents).
Run on
Auto uses WebGPU (full model, ~326 MB) when your browser has it, otherwise the CPU (8-bit model, ~92 MB).
Output
One file per line turns each non-empty line into its own WAV (empty lines are skipped), e.g. one per sentence or word.

Text

0 chars

Audio

Download the model first.

Generated speech appears here. Nothing is uploaded: text and audio stay on this device.

What this tool does

This text to speech tool reads English text aloud with natural sounding AI voices and lets you download the result as a WAV file. It uses Kokoro-82M, an open text-to-speech model with 82 million parameters by hexgrad, through the kokoro-js library and Transformers.js. Everything happens inside your browser: there is no account, no character quota and no server that sees what you type.

Because the model is small, it can run on an ordinary laptop. On browsers with WebGPU (recent Chrome and Edge) it uses the graphics card with the full-precision model; elsewhere it runs on the CPU with an 8-bit quantized version that is smaller and still sounds clear.

How to use it

  1. Click Download model and start. The first time, the model (about 92 MB on CPU, about 326 MB with WebGPU) is downloaded and cached by your browser. A progress bar shows how much is left. Next time it loads from the cache in seconds.
  2. Type or paste your text, or click Sample to try a short paragraph. The character count and the estimated audio length update as you type.
  3. Pick a Voice and a Speed between 0.5× and 2×.
  4. Click Generate speech. Longer text is split into sentence-sized chunks and spoken one chunk at a time, with a “Chunk 3 / 12” counter. Press Stop to end early and keep the audio generated so far.
  5. Listen in the player, then click Download WAV to save a 24 kHz mono file.

To get separate files, set Output to One file per line: each non-empty line becomes its own numbered WAV (01-…, 02-…), and Download all (ZIP) saves them together.

The Share button copies a link with your text, voice and speed, so someone else can generate the same speech in their own browser.

Voices and languages

Kokoro’s browser voices speak English only. There are 20 American English voices (11 female, 9 male) and 8 British English voices (4 female, 4 male). Each voice in the list shows the overall grade from the model card; Heart and Bella (American, female) are rated highest, while Emma (British, female) and Fenrir, Michael and Puck (American, male) are good alternatives.

Pronunciation comes from the espeak-ng phonemizer, so names, acronyms and numbers are usually read correctly, but unusual words may need rephrasing. If most of your text is in another script, the tool warns you, since those characters cannot be voiced.

Example

Load the sample paragraph, keep the default voice Heart and press Generate. The two lines become two chunks, and in a few seconds you get roughly twenty seconds of speech with a short pause between the paragraphs. Switch to George for a British male narrator, or set the speed to 1.3× for a brisker read, and generate again.

Requirements and privacy

The tool needs a modern browser with WebAssembly and Web Workers: current Chrome, Edge, Firefox or Safari on desktop or mobile. WebGPU is optional and only speeds things up. Low-memory phones may struggle with long texts; choose CPU only if WebGPU fails. The model files come from Hugging Face and the ONNX Runtime engine from a public CDN; each voice file (about 0.5 MB) is fetched the first time you use it. Once cached, generation works offline.

FAQ

› Is my text uploaded to a server?

No. The speech is generated on your own device by a model running in your browser. The page downloads the model files once from Hugging Face, but your text and the audio never leave the page.

› Which languages can it speak?

English only. Kokoro-82M v1.0 as packaged for the browser ships 28 voices, 20 with an American accent and 8 with a British accent. Text in other languages or scripts, such as Chinese, Arabic or Russian, will be skipped or mispronounced.

› How big is the download, and do I need it every time?

About 92 MB for the 8-bit model that runs on the CPU, or about 326 MB for the full-precision model used with WebGPU, plus the ONNX Runtime engine and a small file per voice. The browser caches everything, so later visits start in a few seconds without downloading again.

› Can I use the generated audio commercially?

The Kokoro-82M weights and the kokoro-js library are released under the Apache 2.0 license, and the model was trained on permissively licensed audio, so the speech you generate can be used in videos, podcasts or apps. Check the model card on Hugging Face for the full terms.

› Why is generation slow on my computer?

Without WebGPU the model runs on the CPU through WebAssembly, which may take a few seconds per sentence on older laptops and phones. Chrome and Edge with WebGPU are much faster. Long texts are split into chunks, so you can watch progress and stop at any time.