Turn a PDF into speech you can listen to

How to convert PDF to Audio

1. Add your PDF. Drag the file onto the drop zone. The text is extracted locally and shown so you can check it before listening.
2. Press play to listen now. Instant playback uses the voices already built into your device, so it starts immediately with nothing to download.
3. Or generate a downloadable file. Switch to download mode for a higher-quality neural voice. The voice model is fetched the first time you use it and then cached for later.
4. Save the audio. Download the result as WAV or MP3 and put it on your phone, in a podcast app, or anywhere else.

Two modes, and why there are two

Listening immediately and producing a file you can keep are different problems, and one approach cannot do both well.

Instant playback uses the speech engine already built into your operating system — the same voices your device uses for accessibility. It starts the moment you press play with nothing to download, and it handles a 300-page book without difficulty. The catch is that browsers deliberately do not allow this audio to be captured to a file, so it can be listened to but not saved.

Download mode runs a compact neural text-to-speech model directly in your browser instead. The voices are markedly more natural, and because your own machine generates the audio, the result can be written to a real file. The trade-off is a one-time download of roughly 80MB of model weights, fetched only when you first choose this mode and cached afterwards. It is never loaded just because you opened the page.

Genuinely local text-to-speech

Almost every other PDF-to-audio service sends your document to a cloud speech API — Google, Amazon, ElevenLabs or similar. That means the full text of whatever you are listening to is transmitted to a third party, and it usually means an account, a quota and a monthly bill.

Neither mode here does that. Instant playback uses voices already on your device. Download mode runs the model in your browser using WebGPU where available and WebAssembly otherwise. There is no API key to obtain, no per-character charge, and no upload of the document at any point.

That combination is what makes it reasonable to listen to a confidential report on a commute, which is exactly when people want this and exactly when uploading it would be a problem.

What to expect from the output

The audio follows the extracted text, so the quality of extraction sets the ceiling. Flowing prose reads beautifully. A few things predictably do not:

• Tables get read cell by cell, which rarely makes sense aloud. Skip those page ranges.
• Headers, footers and page numbers repeating on every page are detected and dropped so they do not interrupt the flow.
• Formulas, code blocks and long reference lists are read literally and are usually worth excluding.
• Scanned PDFs produce nothing at all, because there is no text to read. Run OCR first.

Frequently asked questions