How ReadLoud works

The parts that matter when you are listening: how fast audio starts, why the highlight stays on the right word, what happens with the screen off, and what you are charged for.

1. Import

Paste text, drop a link, or upload a file. Links go through the same article extraction that powers reader views, so navigation, ads and comment sections are left behind. PDFs are parsed for their text layer and cleaned of running headers, footers and page numbers. EPUB, DOCX, HTML and Markdown keep their headings, lists and quotes.

The text is then split into words, sentences and blocks. Everything downstream, from highlighting to seeking to saving your place, is addressed by word number, which is why your position survives a change of voice or a change of device.

2. Skipping the noise

Before anything is spoken, auto-skip rules decide which words are silent: URLs, email addresses, citation markers like [12] or (Smith, 2020), bracketed asides, and layout noise. Skipped words stay on the page, dimmed, so you always see exactly what was left out. Each rule can be turned off, and a pronunciation dictionary lets you fix names and acronyms once for every document.

3. Synthesis, as you listen

Nothing is generated until you press play. The document is cut into passages of about thirty seconds, with the first three deliberately short so the first sentence plays about two seconds after you press play. Passages are synthesised a little ahead of where you are listening, faster when you listen faster, and generation pauses when you pause. Every passage is cached with the document, so replaying is free.

Four engines are available: Gemini 2.5 Flash (fast, the default), Gemini 2.5 Pro (richest delivery), Gemini 3.1 (preview) and Chirp 3 HD, which we call Narrator. The Gemini engines take a short delivery instruction; Narrator has one even, audiobook-like read in more than fifty languages.

4. Highlighting that is measured, not guessed

Most readers estimate word timing from letter counts and drift after a few paragraphs. ReadLoud measures it twice. As each passage streams in, an energy model aligns the words to the audio in real time, good to about a quarter of a second. As soon as the passage is complete, it is sent to a speech recogniser and the word times are refined against what was actually said, which brings the median error under the recogniser's own resolution. Cross-passage drift is zero by construction, because passage lengths are measured from the decoded audio the browser plays.

If your ear and your eye still disagree by a few milliseconds, there is a sync slider in settings.

5. Playing on a phone with the screen off

A whole document is delivered as one continuous audio stream, so the phone never has to switch tracks in the background, which is the thing that usually stops playback when the screen locks. Lock-screen controls (play, pause, skip, previous and next sentence, scrub) come from the Media Session API. Safari gets the same audio as an HLS playlist, which is what Apple's player was built for. Speed changes travel on a side channel so the stream never reloads. A dropped connection is detected and reopened at the last good position without you doing anything.

6. What you pay for

Your balance is in minutes of generated audio. When a passage finishes synthesising, its real length is deducted. Cached audio is never charged again, and switching voice generates new audio, which is. Listening at 2x costs the same as at 1x because the audio is the same length. When the balance runs out, playback stops at the end of the current passage and you can add minutes without losing your place.

7. Your library

Every document keeps its progress to the sentence, on every device you sign in on. Finished audio can be downloaded as a single MP3. Deleting a document removes its text, its audio and its progress.

Try it with a paragraph of your own, one minute free, no account needed.