Questions that arrive.

Yes. Yappy is open source under the MIT license. Being able to hear what you read is an accessibility matter, and accessibility tools should not ask for a credit card. There is no paid tier, no unlockable limits, no ads.

There is no catch: there are no servers to pay for. Synthesis runs on your device, so Yappy has no per-user cost. It's a personal project by its author, published because it already existed and someone else might need it.

Yappy's code is MIT and puts no conditions on you. The voice model (Supertonic 3, by Supertone) has its own open license: check it on the model's page before commercial use, because that part doesn't depend on Yappy.

No. Synthesis, transcription and article extraction happen on your device. The only network Yappy touches is the initial model download and, if you share a URL, the visit to that URL from your own phone or computer. No accounts, no telemetry, no analytics.

It's a direct connection between your two devices, encrypted end to end (QUIC, using iroh's protocol). Pairing happens by pointing the camera at a QR code your computer shows; the resulting key can be revoked from the computer at any time. There is no Yappy server in the middle, because no Yappy server exists.

Yes, forever. After the initial model download (~380 MB) you can use Yappy in airplane mode: paste text, open documents, synthesize, export. Only sharing new URLs needs a network, for obvious reasons.

Supertonic 3, an open model by Supertone of about 66 million parameters that fits in 380 MB and synthesizes faster than real time on any laptop from the last few years. It works at the character level: that's why the guionizador exists, turning numbers, acronyms and symbols into words before the voice sees them.

Arabic, Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hindi, Hungarian, Indonesian, Italian, Japanese, Korean, Latvian, Lithuanian, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Turkish, Ukrainian and Vietnamese. All ten voices are polyglot: the same voice reads any of the 31 languages.

Yappy detects the document's language and, paragraph by paragraph, each piece's: an English quote inside a Spanish essay is read in English, with the same voice, without you touching anything. And if detection gets it wrong, you can set the language by hand.

The model loads into memory the first time. It stays warm after that, and the following sentences come out on the fly.

Web articles (the page cleans itself), YouTube videos (via their transcript), PDF, EPUB, Word, ODT, Markdown, plain text and voice notes. On mobile, everything through the share sheet; on desktop, also by dragging or with the shortcuts.

On desktop they go through local OCR (PaddleOCR) before being read: nothing leaves your computer there either. A crooked or blurry scan will give what any OCR can give, but a good one reads through whole.

From the library itself: every document has its forget option. Exported files (.m4b, .mp3, .wav) are yours and live wherever you saved them; deleting or moving them is a job for Finder or Files.

System readers read the screen as-is, with the system's prosody and no memory. Yappy prepares the text (cleans the article, verbalizes numbers and acronyms, marks the pauses), reads it with a natural voice, remembers where you were, exports audiobooks and keeps playing with the screen locked. Cousin tools, different trades.

In the GitHub issues. With the text that failed (if you can share it) and your system, it gets fixed sooner.

Since the model reads characters, the quick path is writing the word as it sounds in the editor. If it's a general pattern (an abbreviation, an acronym, a unit), tell it on GitHub: the guionizador's tables grow from those cases.