Credits and licenses

PageVoice runs inference on your device. Only fixed model and runtime files are downloaded. Your book text is never part of those requests. The optional device-speech engine uses your operating system; its privacy and offline availability depend on that platform.

Speech is generated. Voices do not imply the participation or endorsement of a real speaker. Use audio lawfully and respect the restrictions attached to its model and the copyright of the book.

Models and voices

Component	License and source
Kokoro English, US and UK voice styles	Apache-2.0. hexgrad, ONNX conversion. License. English phonemization only in this app.
Piper DaveFX, Spain Spanish	MIT voice repository; CC0 speech dataset. Model card, Pinned weights and source.
Piper Sharvard, two Spain Spanish speakers	MIT voice repository; CC-BY-3.0 corpus. Attribution: Vincent Aubanel, Maria Luisa García Lecumberri and Martin Cooke, Sharvard corpus. Model card. Speaker gender/age is not independently verified.
Supertonic 2, F1 and M1	OpenRAIL-M, including use restrictions. Supertone weights. Upstream archived September 2026. Experimental comparison; not a measured quality improvement.
Multilingual MiniLM meaning search	Apache-2.0. Sentence Transformers, Xenova ONNX conversion. Matches are inferred, with source citations.
Multilingual name detection	AFL-3.0. Davlan, Xenova conversion. Trained on news; names and speakers in fiction can be wrong.
English and Spanish OCR data	Apache-2.0. Tesseract tessdata, integer LSTM data distributed by @tesseract.js-data/eng and /spa 1.0.0. License.

Runtimes and redistribution

Model licenses and library licenses are separate. PageVoice's original code is MIT; downloaded third-party software retains its own license.

kokoro-js 1.2.1 and phonemizer 1.2.1: Apache-2.0 wrappers. The compiled eSpeak NG component is GPL-3.0. Phonemizer source/build instructions, eSpeak NG source.
@diffusionstudio/piper-wasm 1.0.0: MIT wrapper. Piper phonemizer and eSpeak NG include GPL-3.0 components. Wrapper source, Piper phonemizer source. PageVoice adds an ES-module export shim; the build script records its exact bytes and hash.
ONNX Runtime Web 1.30.0, plus Transformers.js's pinned 1.22.0-dev runtime: MIT. Source. Transformers.js 3.8.1: Apache-2.0.
pdf.js 6.3.289 and Tesseract.js 7.0.0: Apache-2.0, PDF source, OCR source.
@breezystack/lamejs 1.2.7: LGPL-3.0. Source. MP3 export runs in a worker.
@ffmpeg/ffmpeg 0.12.15: MIT wrapper. @ffmpeg/core 0.12.10 includes GPL-2.0-or-later and LGPL components. Source, build recipes and dependency list, FFmpeg fork source. PageVoice uses the single-thread build without binary modification.
React / ReactDOM 19.3.0, xmldom 0.9.12: MIT. JSZip 3.10.2: MIT or GPL-3.0; PageVoice uses it under MIT. React notice, xmldom notice, JSZip notice.
Newsreader and Source Sans 3, self-hosted Latin/Spanish subsets: SIL OFL-1.1. Newsreader notice, Source Sans notice.

Exact model revisions, file sizes and SHA-256 hashes are checked into web/src/offline/model-assets.json. Runtime versions are pinned in the npm lockfile and hashes in runtime-assets.json. No models download automatically.

Back to PageVoice