Kokoro Neural Voices for NVDA

Natural, fully offline neural text-to-speech for the NVDA screen reader, powered by the Kokoro voices. Fifty-plus voices across nine languages, voice blending, selectable model quality, and an audio cache that makes typing echo and navigation instant.

• American and British English, Spanish, French, Hindi, Italian, Brazilian Portuguese, Japanese, and Mandarin Chinese, with automatic language switching.
• Runs entirely on your machine. No internet after the model is downloaded, no accounts, no telemetry.
• A bundled 64-bit inference helper keeps heavy code out of NVDA, so a model fault can never crash your screen reader, and one build works on both 32-bit NVDA (2025.x) and 64-bit NVDA (2026.1+).

Quick start

1. Install the add-on (NVDA menu, Tools, Add-on Store, or Install from external source, then choose the .nvda-addon file).
2. Open NVDA menu, Preferences, Settings, Speech, and choose "Kokoro Neural Voices" as your synthesizer.
3. The first time you select it, accept the prompt to download the voice model and voices (about 190 MB for the balanced model plus voices).

Full instructions are in docs/USER_GUIDE.md.

Performance note

Kokoro is a distinctive, higher-quality neural model, but it is heavier than some alternatives. On a mid-range CPU new prose starts in roughly half a second to a second, while character echo and common navigation words are made instant by an audio cache that is warmed during idle time and saved between sessions. Cancel is effectively instant. If you want the fastest possible neural voices on CPU, see the companion Piper add-on; a measured comparison is in that project's COMPARISON.md.

Documentation

• User guide - install, first-run download, settings, troubleshooting.
• Architecture - the helper, driver, cache, and audio pipeline.
• IPC protocol - the driver-to-helper message reference.
• Building from source.
• Testing and verification.
• Add-on Store submission.
• Contributing and coding standards.
• Changelog, licensing, third-party notices.
• The original product requirements: PRD.md; the implementation plan: docs/plans/.

Project layout

koko/
  addon/            NVDA add-on tree (packaged into the .nvda-addon)
    synthDrivers/kokoro/     the SynthDriver and its helper-facing modules
    globalPlugins/kokoroManager/   Tools-menu entry for the component manager
    manifest.ini, installTasks.py, doc/
  helper/           Rust crate: kokoro-helper.exe (ONNX inference + espeak-ng)
  tools/            build.py (packaging), fetch_assets.py, benchmarks
  tests/python/     pytest suite with NVDA API stubs
  docs/             this documentation set, plus the implementation plan
  PRD.md            product requirements

Credits

The model and voices come from the kokoro-onnx project (Kokoro-82M, Apache-2.0). Phonemization uses espeak-ng. Inference uses ONNX Runtime via the ort crate. Not affiliated with NV Access.