Papervoice
Turns the PDFs you own into audiobooks. Entirely on your phone — no account, no
subscription, and nothing you read is ever uploaded.
Import a PDF, press play, and follow along as each sentence is highlighted. The
voices are neural text-to-speech running on the device itself, so it works with
the aeroplane mode on.
Android, GPLv3. Free, and free of ads, trackers and accounts.
Get it on Google Play · papervoice.app
Install
From Google Play,
like any other app.
For a phone without Play, every release also carries the APK under one stable
name, so this link always gets the latest:
papervoice.apk.
Android will warn you about installing from outside Google Play — that is
expected. Each release has a papervoice.apk.sha256 beside it to check the
file against first:
sha256sum -c papervoice.apk.sha256
For updates without re-downloading by hand, point
Obtainium at this repository.
What it does
- PDF to speech. Text is extracted with pdf.js, split into sentences, and
read aloud by a neural voice.
- Follow along. The spoken sentence is highlighted and the page scrolls with
it. Tap any sentence to jump there.
- Fifteen natural voices, eight male and seven female, three British. Every
one is public domain — see docs/voice-credits.md,
which is generated from an audit rather than written by hand.
- Your phone's own voices work too, with nothing to download, listed by
name alongside the natural ones — American, British, Australian and Indian.
- A voice for the book. It suggests voices from the book's own text: a
story or a textbook, British or American spelling.
- It remembers where you were, per book, to the sentence.
- Sleep timer: stop after a set time, or at the end of the chapter.
- Groups, so a shelf of textbooks is not one long list.
- Paper, light and dark themes; adjustable text size; speed from 0.5× to 3×.
It sends nothing on its own. The only things that ever leave the phone are a
voice you choose to download, and an optional questionnaire or feedback when
you press Send — never anything from your documents. The
privacy policy lists exactly what each one
contains.
What it does not do
- Scanned PDFs. There is no OCR. A PDF of photographs of pages has no text
to read, and the app says so rather than pretending.
- DRM-protected books. Nothing from Kindle or Adobe DE.
- iOS. Android only.
Building it
Requires Node 20+, a JDK, and the Android SDK.
npm install
npm run android # debug build on a connected device or emulator
npm run apk # release APK, both ABIs
npm run apk -- --fast # arm64 only, for testing a release build quickly
npm test # no device needed
Build settings that should not be committed go in .env and .env.local
(both git-ignored), which every build script loads itself:
EXPO_PUBLIC_FEEDBACK_ENDPOINT for where feedback is sent, and
GRADLE_USER_HOME if Gradle should keep its cache somewhere other than the
default.
npm run release -- --publish builds, signs and publishes a GitHub release,
as a draft first, with the stable-named APK and checksums beside the
versioned one.
android/ is generated by expo prebuild and is not committed. Anything
that needs to change in it lives in plugins/ as a config plugin, because a
hand edit there disappears at the next prebuild with no sign that it has.
The emulator
npm run emu:create # once
npm run emu # boot it
npm run emu -- --wipe # boot it as a phone nobody has opened
npm run apk -- --emu # x86_64 build, the only kind an emulator installs
Layout, navigation, insets and empty states are faithful on an emulator.
Voice speed, audio routing and battery are not — those need a real phone,
because the emulator runs on your computer's cores with no thermal limit.
The website
papervoice.app is the static page in site/,
published by GitHub Actions. site/README.md covers the voice
samples, the screenshots, the demo video and the release details written in at
deploy.
Contributing
Start with docs/ARCHITECTURE.md, which explains how a
PDF becomes sound and where each part of that lives.
The code is commented more heavily than is fashionable, and the comments are
mostly about why rather than what. Where something looks odd there is
usually a paragraph saying which bug it came from. Please keep that up: a
comment explaining what a line does is noise, one explaining why the obvious
version was wrong is the valuable part.
Some things worth knowing before you change anything:
npm test needs no device. All the logic that can be tested without one
is, and there are fixtures for the extraction pipeline.- Licences are enforced by a test.
npm run audit:licences re-derives every
voice's licence from its source and follows "finetuned from X" chains;
__tests__/licences.test.ts fails if the catalogue drifts from what was
verified. Adding a voice means running the audit.
- Check a model before adopting it:
npm run audit:licences -- check <model>
judges a HuggingFace repo or Piper voice and says whether it can be shipped.
- Measure, don't reason. Several bugs here survived because someone
reasoned about what the code must be doing instead of running it. There are
tools for the recurring questions —
npm run phonemes compares what a
phonemiser produces against what a model accepts.
Why it is GPLv3
Not by choice. eSpeak NG does the letters-to-sounds step for every neural voice,
it is GPLv3, and it is compiled into the native library this app links.
That has a consequence worth stating plainly: this app cannot be sold as
closed-source software, and the natural voices cannot be moved into a
proprietary build without replacing that one component. Everything else here is
permissively licensed.
LICENSE-EXCEPTION adds the section 7 permission that GPL
apps need to be distributed through an app store, conditional on the
unencumbered source remaining available somewhere anyone can reach without an
account.
Credits
The voices are other people's work, released freely. They are listed with the
page each licence was read from in
docs/voice-credits.md — in particular Bryce Beattie,
who trained most of them from public-domain LibriVox recordings and wrote down
the lineage, which is rarer than it should be.
Built on sherpa-onnx,
Piper, pdf.js
and Expo.
