The voices are Kokoro, an open AI speech model that sounds far more natural than the robotic voices built into computers and phones, and that until now meant a paid online service with a character limit. Here it runs in your browser: type or paste text in English, Spanish, Italian, Brazilian Portuguese, Hindi, Japanese or Mandarin Chinese, pick one of 48 voices, and listen or save the result as an MP3. There is no sign-up, no limit on how much you convert, and the text never leaves your computer. It is free to use, with no account and no upload limit, and it runs entirely in your browser: nothing is uploaded, and while the page stays open, a feature you have already used keeps working offline.
Before a voice can read, every word has to become sounds. This page does that itself with the same American and British pronunciation dictionaries Kokoro was trained with, over 180,000 words each. It writes out numbers, money, dates and times as they are said, spells acronyms like FBI letter by letter, and guesses the rest with a small neural model. A word read wrongly can be respelled. Spanish, Italian, Portuguese and Hindi have spelling rules regular enough to work from directly, so for them the page uses rules instead of a dictionary: syllables, stress, silent letters, nasal vowels, Italian double consonants, the Hindi vowel that is written but often not said, and the sounds that change with their neighbours. Japanese and Chinese cannot be read from the characters alone, so for them the page downloads a dictionary the first time: it splits the text into words and looks up how each is said, in kana for Japanese and in pinyin with its tones for Chinese. Each language reads numbers, money, percentages and times its own way; Hindi counts in lakhs and crores.
A report to get through on the commute, lecture notes to go over on a walk, a novel without an audiobook: open the PDF and listen. Open the page with the document box first, or use the box above the text.
- Can I use the audio in a video or a podcast?
- Yes. Kokoro's model and voices are released under the Apache 2.0 licence, which allows any use, including commercial, and this page adds no conditions of its own.
- Which languages does it speak?
- English, in American and British accents; Spanish, pronounced as in Spain (the c of "cielo" and the z of "zapato" said as th); Italian; Portuguese as spoken in Brazil; Hindi, written in Devanagari; Japanese; and Mandarin Chinese, written in simplified characters. There are two Italian voices, three Brazilian, four Hindi, one Japanese and eight Chinese ones.
- Are the voices in other languages as good as the English ones?
- Not quite. Kokoro was trained on much less speech in these languages - for Italian, less than an hour - so they can sound flatter, and a rare word or name may come out wrong. The Chinese voices are the weakest by Kokoro's own grades. A respelling fixes a word that is read wrongly.
- Can it read a PDF, an e-book or a Word document aloud?
- Yes. Open a PDF, an EPUB without DRM, a Word .docx or a text file, and the page splits it into pages (PDF), chapters (EPUB, and Word documents with headings) or parts, so you can pick where to start and stop. It is read as it is made, or saved as an MP3 - one file, or one per chapter in a ZIP - written as it goes, so even a whole book does not fill your memory. A book takes a while: on most computers the voice is made about as fast as it plays, much faster with the graphics card. A scanned PDF has pictures of pages and no text, so run it through the OCR Toolkit first.
- Why does it say my PDF has no text to read?
- The PDF is a scan: each page is a picture of the paper, with no text inside for a voice to read. Run it through the OCR Toolkit first, which recognises the words and saves a searchable PDF, then open that here.
- Are a PDF's headers, footers and page numbers read out?
- Whatever text is on the page is read, in the order the PDF stores it, so running headers and page numbers can be read too, and a page laid out in columns or with many boxes may come out in an odd order. Picking the pages you want avoids covers, contents and indexes. EPUB and Word files have no page furniture, so they read more cleanly.
- How big is the MP3 of a PDF or book?
- Speech from a document is saved at 96 kb/s, about 43 MB an hour, which is plenty for a voice. A long book comes out as a large single file, so one MP3 per chapter can be easier to handle and to resume on a phone.
- Why does Japanese or Chinese download a dictionary?
- Written Japanese and Chinese do not show how a character is said: the same character has different readings in different words, and there are no spaces between words. So the first time you choose Japanese, the page downloads a Japanese dictionary of about 18 MB (kuromoji with IPADIC); for Chinese, a word list and pinyin dictionary of about 5 MB (jieba and pinyin-pro). Your browser keeps them, and nothing is downloaded until you pick the language. If a name or an unusual word is read wrongly, add a respelling in kana or in other characters that have the sound you want.
- Why is the first time slow?
- The voice model, about 92 MB, is downloaded once and kept by your browser. After that it starts in a few seconds, and on most computers makes speech about as fast as it plays. Where the browser can use the graphics card (WebGPU), ticking Use the graphics card switches to the full model - a 326 MB download - which is many times faster.
- Does the text I type leave my computer?
- No. The model is downloaded from huggingface.co and the dictionaries from this site, and everything is spoken on your own computer. A PDF or document you open is read in this browser tab too; nothing is uploaded.