Lore Foundry TTS Pack
The on-device "Natural voices" model pack for Lore Foundry — downloaded once by the app, after which all narration runs entirely on the user's machine. Nothing is sent anywhere at synthesis time.
Contents
| Path | What | Source / license |
|---|---|---|
kokoro/kokoro_with_dur_uint8.onnx |
Kokoro-82M TTS, static-uint8 quantized, with a pred_dur output exposing per-phoneme durations (drives word-accurate read-along highlighting) |
Graph: onnx-community/Kokoro-82M-v1.0-ONNX (Apache-2.0), model: hexgrad/Kokoro-82M (Apache-2.0). The pred_dur output was added by a one-node graph modification — see scripts/graft_pred_dur.py (the waveform path is untouched; output audio is bit-identical to the source export). |
kokoro/voices/*.bin |
Voice style vectors (55 voices) | hexgrad/Kokoro-82M (Apache-2.0) |
kokoro/config.json, kokoro/tokenizer.json |
Model config + phoneme tokenizer | as above |
openphonemizer/model.onnx, openphonemizer/tokenizer.json |
OpenPhonemizer — out-of-vocabulary grapheme→phoneme fallback (invented/fantasy names) | openphonemizer (BSD-3-Clause-Clear) |
scripts/graft_pred_dur.py |
The auditable graph-surgery script that produced the dur-exposing model | Lore Foundry (Apache-2.0) |
Integrity
Consumers should verify per-file SHA-256 checksums (the Lore Foundry app pins
them). Model file: kokoro_with_dur_uint8.onnx =
3d2cdc1221c91c7932adb9d2fd6a94b1f2918f8a97a7d5417295dcd4d4528204.
Versioning
Pack revisions are tagged; the app pins a revision. Current: v1.