Turn a document into speech you can play anywhere

There is a large category of documents you need to have read but do not need to sit and read: industry reports, long email attachments, papers you are keeping up with, drafts you want to hear for rhythm. Speech turns those into something you can process while doing something else.

How it works and why that matters

The audio export tool extracts the document's text and runs a speech model on your own hardware, producing a real WAV file rather than a stream you have to stay connected for.

The on-device part is the substantive difference from the usual options. Cloud text-to-speech services require the document text to be sent to a provider, and the documents people most want read aloud — internal reports, legal papers, medical letters, unpublished drafts — are precisely the ones least suited to being posted to an API. Local synthesis removes that trade-off entirely.

The trade-offs it introduces are honest ones: the first run downloads a voice model, generation takes real time proportional to the document length, and the voice is good rather than indistinguishable from a human narrator.

The text matters more than the voice

This is the part people skip, and it is the difference between a listenable file and one you abandon after ninety seconds.

A PDF's text layer contains everything drawn on the page, in drawing order. Read aloud verbatim, that includes:

• Running headers and footers, so every page interrupts the sentence with the document title and a page number.
• Footnote markers, which become stray digits mid-sentence.
• Footnote text, read in whatever position it was drawn.
• Tables, which become a stream of unrelated numbers.
• Figure captions and axis labels, dropped into the middle of paragraphs.
• Reference lists, which are excruciating as audio.
• Hyphenated line breaks, producing pronunciations like "manage — ment".

So the workflow that produces something worth listening to is:

1. Extract the text with the text extraction tool and look at it.
2. Delete page furniture, reference sections, and tables.
3. Rejoin hyphenated words and paragraph line breaks.
4. Expand abbreviations that will be mispronounced, and spell out anything where the pronunciation matters.
5. Then synthesise.

For a long document, ten minutes of cleanup saves an hour of confusion. The extraction guide covers the cleanup patterns in more detail.

Practical generation tips

• Test on two pages first. Generate a short sample, listen to it, and adjust before committing to a 200-page document.
• Split long documents by chapter. Separate files are easier to navigate, easier to resume, and easier to regenerate if one section needs fixing.
• Check numbers and acronyms. These are where synthesis most often diverges from what you meant.
• Expect real processing time. A long document takes a while. Start it and do something else.
• Keep the file size in mind. Uncompressed audio is large; a full-length book produces a substantial file.

What audio is good for, and what it is not

Good for: getting through material during a commute, cooking, or walking; first passes over long documents to decide what deserves proper reading; hearing your own drafts, which exposes clumsy sentences that the eye skips; accessibility for readers who find extended screen reading difficult; and revisiting familiar material.

Not good for: studying something technical, where you need to re-read a sentence three times; anything where tables, equations, or figures carry the argument; documents you need to annotate; and reference material you dip into rather than read through.

Audio is linear. You cannot skim it, and glancing back at the previous paragraph means scrubbing a timeline. For dense material, reading remains faster.

Related options

If the actual problem is that a PDF is uncomfortable to read on a phone rather than that you want to listen, converting to a reflowable format is the better fix — the EPUB guide covers that, and most reading apps have their own speech function built in, which gives you both options from one file.