richardr1126/Kokoro-82M-OpenReader

Kokoro-82M for OpenReader

Kokoro-82M v1.0 (hexgrad/Kokoro-82M at f3ff3571791e39611d31c381e3a41a3af07b4987), converted for the OpenReader apps. Both conversions come from the same checkpoint, and each app pins a revision of this repository.

Path	Used by	Contents
*.aimodel/, runtime.json, config.json, voices/*.npy	OpenReader iOS	Core AI bundles (predictor, prosody, vocoder) and voice packs
onnx/	OpenReader compute worker (kokoro-local)	model.onnx with per-token durations, raw voice packs; see onnx/README.md
g2p/	Both	English letter-to-sound rules learned from Misaki's dictionaries (Apache-2.0), and Spanish, Hindi, Italian and Brazilian Portuguese G2P tables learned from Kokoro's own readings (CC BY 2.0 FR); see g2p/README.md

This repository was named richardr1126/Kokoro-82M-CoreAI; links using the old name redirect here.

Core AI

Reproduce the dynamic-length Kokoro export

Use Python 3.11 on macOS 27 and requirements.txt, with Apple's package index. Install en_core_web_sm 3.8.0 for Misaki's official English reference frontend. The source revisions and export license are recorded in PROVENANCE.md.

python tools/kokoro/export_coreai.py --out-dir Models/kokoro-work/export
USE_OS_COREAI=1 python tools/kokoro/export_coreai.py --out-dir Models/kokoro-work/export --verify --voice af_heart --text "A long reference passage goes here."

The reference is PyTorch, never generated from Core AI's answer. The script checks magnitude-spectrum correlation (>=0.99) and reports waveform correlation; it writes a WAV without playing it. Lengths are dynamic up to 512 tokens and 2048 frames (--max-tokens, --max-frames); --verify --pad checks a pass padded to those bounds instead of at its own length.

Checkpoints, environments and audio stay in ignored Models/. Publish only the three aimodel bundles, runtime.json, the pinned config.json, source/recipe and notices. Existing voice NPYs remain unchanged. Validate the actual Swift chain on hardware: copy the files into the app container's Library/Application Support/KokoroModel/, one full destination path per file, then run Scripts/check-iphone-tts.py --engine kokoro --case muted. The simulator lacks Core AI and cannot validate these bundles.