| license: other | |
| tags: [coreml, singing, lyrics-alignment, source-separation] | |
| # SingQ models | |
| Core ML models downloaded on first use by **SingQ**, a macOS app for singing practice. | |
| | File | Use | Source | License | | |
| | --- | --- | --- | --- | | |
| | `wav2vec2-ko.mlpackage.zip` | Korean lyric alignment (CTC) | [kresnik/wav2vec2-large-xlsr-korean](https://huggingface.co/kresnik/wav2vec2-large-xlsr-korean), trained on [Zeroth-Korean](https://openslr.org/40/) (CC BY 4.0) | Apache-2.0 | | |
| | `wav2vec2-en.mlpackage.zip` | English lyric alignment (CTC) | [facebook/wav2vec2-large-960h-lv60-self](https://huggingface.co/facebook/wav2vec2-large-960h-lv60-self) | Apache-2.0 | | |
| | `htdemucs.mlpackage.zip` | Vocal / accompaniment separation | [Hybrid Transformer Demucs](https://github.com/facebookresearch/demucs) by Meta AI; Core ML conversion from [slaughters85j/htdemucs-coreml](https://huggingface.co/slaughters85j/htdemucs-coreml) | MIT | | |
| The wav2vec2 models were converted to Core ML (English FP16, Korean FP32 because the XLS-R model overflows in FP16; fixed 34-second input at 16 kHz, log-softmax output every 20 ms) without retraining. | |
| `models.json` lists each file's size and SHA-256. | |