Kokoro TTS Online

• Real Kokoro-82M — the open-weight model, not a lookalike — via the onnx-community ONNX build
• 28 verified voices across American/British English accents, male and female
• Runs locally in your browser: one-time model download, then unlimited offline-capable generation
• No API key, no signup, no character limits, no watermarked audio
• Private by architecture — scripts and drafts never leave your device

Frequently Asked Questions

Is this the real Kokoro model or an imitation?

The real one: the tool loads onnx-community/Kokoro-82M-v1.0-ONNX, the community ONNX export of the open-weight Kokoro-82M model, and runs it with transformers.js in your browser. The 28 voices are Kokoro's own voice embeddings.

Why is Kokoro TTS free here — what's the catch?

Architecture, not generosity: the model executes on YOUR hardware, so there's no server GPU bill to pass on. The trade-offs are the one-time ~80 MB download and generation speed that depends on your device. No account, no caps, no watermark.

Can I use the generated audio commercially?

Kokoro-82M's weights are released under the Apache 2.0 license, which permits commercial use of the model and its output. As always, you're responsible for the content of the speech itself (no impersonation or deceptive use).

Does it work offline?

After the first generation caches the model, synthesis itself runs locally — reload while offline and generation still works in most browsers as long as the cache persists. The initial model download requires a connection.

Why was my first generation slow?

The first run downloads the ~80 MB ONNX model and initializes the runtime. Subsequent generations skip both steps. Speed after warm-up depends on hardware — WebGPU-capable browsers are dramatically faster than the WebAssembly fallback.

What languages does Kokoro support?

This build focuses on English, with American and British accent voices (the af/am/bf/bm voice prefixes). For other languages, a hosted multi-language TTS service is currently the better tool — that's the honest limit of the 82M model's voice set here.