The AAC reference app chooses which of a device's own voices may speak for a
person (access-core, src/voice.rs). Part of what it goes by is what Readium
Speech documents about the voices browsers expose: which are novelty or very
low quality voices, how Apple marks a voice's quality in its name, and, for
each language, the voices known and their quality, in Readium's order.

generate.py beside this file reads those facts at one pinned commit and
writes them into crates/access-core/src/voice/readium.rs as Rust tables, so
the app carries them in its own code. Nothing is fetched while the app runs,
and nothing about a person's voices leaves their device.

Readium Speech (@readium/speech v0.13.0), by the Readium Foundation.
  Used: json/<language>.json for the app's languages, json/filters/novelty.json,
  json/filters/veryLowQuality.json and json/localizedNames/apple.json, reduced
  to voice names, quality and order (network voices left out).
  The code is licensed under the BSD 3-Clause License: LICENSE.txt beside
  this file (the project's LICENSE, verbatim at the same commit).
  The voice data began as Hadrien Gardeur's web-speech-recommended-voices,
  released under CC0 1.0 (https://github.com/HadrienGardeur/web-speech-recommended-voices).
  Source: https://github.com/readium/speech (tag v0.13.0, commit
          4a7b73dd37a736e6762497ea7d7bc773c0fe1a30, read 2026-10-04; its json/
          is byte-identical to v0.11.0's), https://www.npmjs.com/package/@readium/speech