Natural text-to-speech synthesized on your device.

A free text-to-speech tool that generates natural-sounding speech entirely in your browser with the Kokoro neural voice model. Use it for narration drafts, video voiceovers, language learning, accessibility, or hearing your own writing read aloud — with no account, no API keys, and no per-word pricing.

Two engines ship in the same interface: 28 American and British English voices with official naturalness grades, and 100 Mandarin voices that read mixed Chinese–English text, for 128 voices in total. Speed control, long-document mode, batch queues and local history round out the studio.

How it works

Kokoro-82M is a compact, high-quality neural TTS model running locally through Transformers.js, with WebGPU acceleration in Chrome or Edge and a WebAssembly path everywhere else. The English engine is the v1.0 build via kokoro-js; the Mandarin engine is the v1.1-zh build with built-in Chinese G2P. Voices and model weights download once and are then cached, after which synthesis works offline. If a WebGPU run fails at runtime, the tool transparently retries once on WASM instead of showing a cryptic error.

128 voices and the voice browser

The voice browser is built for a library this large: full-text search by name or ID, filters for gender and accent (Mandarin, American, British), fixed-sentence previews that reuse cached audio, favorites stored on your device, and a recently-used list — alongside the official A–F naturalness grades the English voices carry. Picking a voice drops it straight back into the synthesizer; there is no queue or server-side catalog call.

Long-document mode

You are not limited to a single short sentence. Text longer than the safe per-call budget is split at sentence boundaries — with a hard fallback for very long sentences — synthesized chunk by chunk with live progress (chunk 3 of 12), and stitched together with short equal-gain crossfades so the joins don't click. The free tier reads documents up to 10,000 characters; Pro goes to 100,000. You can cancel a long run at any point, and an English voice reading CJK text is caught up front with a clear hint to switch to a Mandarin voice.

Formats, batch queues and history

WAV (16-bit, 24 kHz) is always exportable. Chromium browsers additionally expose MP3 via WebCodecs; other browsers use Opus/WebM, and the available formats are detected and labeled honestly rather than failing at click time. Batch mode queues up to 5 texts for free (50 on Pro), can load multiple .txt files at once, runs them serially with per-item progress, and exports the set individually or as a ZIP. The history panel keeps the latest twelve generations — text, voice and speed — right in the browser for one-click regeneration, and can be cleared at any time.

Limits and privacy

Neither tier has a run count, a monthly character quota, or a watermark, and both include commercial use; only per-document length (10,000 vs 100,000 characters) and batch size (5 vs 50 texts) differ. Your text never leaves the browser — scripts, notes and drafts are private by construction, which is the main reason to prefer this over a cloud TTS API for confidential material.

Common questions

Why can English voices read only English? The phoneme and vocabulary coverage belongs to the engine; the Mandarin v1.1 engine is the one designed for Chinese and mixed text, and the tool enforces this instead of producing garbled speech.

Can I generate audio offline? Yes — once a voice and its model are cached, synthesis needs no connection.

Is MP3 guaranteed? It is available wherever WebCodecs MP3 encoding exists (Chromium); WAV is the universal format and Opus/WebM is offered where MP3 is absent.

Is the free tier really free for commercial voiceovers? Yes, with no per-use fee and no watermark — only document length and batch size distinguish Pro.