# Third-party components and licences

My Accent Coach's own code and coaching material are proprietary, all rights reserved (see `LICENSE`).
The components below are included in this repository under their own licences.
**Not everything here may be used commercially: see the Expresso recordings below.**

| Component | Where | Licence | Source | Changes made |
|---|---|---|---|---|
| ONNX Runtime Web 1.31.1 | `vendor/ort*` | MIT (`licenses/MIT-onnxruntime.txt`) | https://github.com/microsoft/onnxruntime | none |
| qrcode-generator 2.0.4 | `vendor/qrcode.js` | MIT (notice in the file header) | https://github.com/kazuhikoarase/qrcode-generator | none |
| misaki 0.7.4 US pronunciation dictionaries (us_gold, us_silver) | `tts/us_*.json` | Apache-2.0 (`licenses/Apache-2.0.txt`, `licenses/Apache-2.0-misaki-NOTICE.txt`) | https://github.com/hexgrad/misaki | none (lookup logic re-implemented in `js/g2p.js`) |
| Kokoro-82M v1.0 voice model and voices | `tts/` | Apache-2.0 (`licenses/Apache-2.0.txt`) | https://huggingface.co/hexgrad/Kokoro-82M (ONNX export: https://github.com/thewh1teagle/kokoro-onnx, MIT) | added an output reporting phoneme durations; weights quantized to 8-bit except the waveform generator; split into two files (`tools/patch_model.py`, `tools/export_web_tts.py`) |
| Whisper base.en speech recognition model (OpenAI), ONNX conversion by onnx-community | `ear/encoder.onnx`, `ear/decoder.part1-3`, `ear/config.json`, `ear/mel.bin` | MIT (`licenses/MIT-whisper.txt`); the conversion is published under Apache-2.0 (`licenses/Apache-2.0.txt`) | https://github.com/openai/whisper, https://huggingface.co/onnx-community/whisper-base.en | 8-bit quantized ONNX files used as published (renamed; decoder with key/value cache, split into 3 parts); tokenizer and mel filter bank re-packaged as `config.json` and `mel.bin`; decoding re-implemented in `js/whisper-worker.js` |
| Reference recordings generated with Kokoro-82M | `audio/` (except `audio/real/` and `audio/readers/`) | Apache-2.0 model; recordings generated for this project | — | — |
| Expresso dataset recordings (Nguyen et al., 2023) | `audio/real/` | **CC BY-NC 4.0: non-commercial use only** (https://creativecommons.org/licenses/by-nc/4.0/) | https://speechbot.github.io/expresso/ | selected clips, trimmed silence, normalized volume, re-encoded as MP3; word timings computed |
| LibriTTS-R recordings (Koizumi et al., 2023), from LibriVox audiobooks; speakers 4446 and 260, test-clean | `audio/readers/` | CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/); commercial use allowed with credit | https://www.openslr.org/141/ | selected clips, trimmed silence, normalized volume, re-encoded as MP3; word timings computed; text punctuation tidied |
| Instrument Sans, IBM Plex Sans, Courier Prime | `fonts/` | SIL Open Font License 1.1 (`licenses/OFL-*.txt`) | Google Fonts / Fontsource | subset files (Latin) only |

## Commercial use of the included recordings
The Expresso recordings in `audio/real/` (and `content/units/real.json`) are non-commercial only and are not part of any paid version. All other components listed above allow commercial use under their licences.
