I built this free AI voice generator to turn text into speech without an account. Paste your text, choose one of 28 American or British English voices, and the audio is generated on your device.
It runs the open-source Kokoro model in your browser. It tries WebGPU when available and falls back to slower WebAssembly when needed.
[0.5s] in the text — change the number to whatever you prefer.Every clip reads the same line: “Some mornings, the coffee tastes better than usual — and I never quite know why.”
Voice quality varies. Training duration matters, but so do the source audio and how well it aligns with the text. Heart has the highest official overall grade.
WAV is uncompressed PCM (24 kHz, 16-bit mono, about 2.9 MB per minute). MP3 uses 128 kbps CBR and is about 960 KB per minute. Kokoro produces 24 kHz speech, so neither download is studio-quality audio.
Yes, but smaller sections are more practical. The 50,000-character limit prevents unusually large jobs; it does not guarantee that a single job of that size will finish on every device.
The generator offers 20 American English and 8 British English voices. Kokoro voices for other languages are not included.
A browser with WebAssembly is required. The generator uses WebGPU when the browser and device expose it. Otherwise, it starts with the slower WASM fallback.
The first generation downloads Kokoro's scripts, tokenizer, model, and selected voice. The WebGPU model is roughly 326 MB. The WASM model is roughly 92 MB, plus its runtime and other assets. A failed WebGPU attempt may download both models. Your browser can cache these files, but it may evict or revalidate them. Clear this site's stored data in your browser settings to remove them.
Speech synthesis runs on your device. The generator does not upload the text you enter or the audio it creates. It still downloads scripts, model files, and voices from jsDelivr and Hugging Face. Those providers receive ordinary request metadata, including your IP address, browser, requested URL, and the time of the request. In production, Fathom records page views and the fixedVoice generated event. It does not receive your text or audio from this generator.
The license labels alone do not provide blanket commercial clearance. The reviewed Kokoro model, ONNX conversion, JavaScript code, and repository-distributed voice tensors are labeled Apache-2.0. That does not settle ownership of generated audio or any other rights that may apply. You are responsible for the text you supply and for checking whether your intended use is permitted.