| --- |
| license: mit |
| tags: |
| - coreml |
| - audio |
| - source-separation |
| --- |
| |
| # Spleeter 2stems for Core ML (vocal mask, 1536 bins, FP16) |
|
|
| A Core ML conversion of [Spleeter](https://github.com/deezer/spleeter)'s pretrained 2stems model (v1.4.0) by Deezer, used to separate voice from music in Clarity Sound Studio, a free, open-source macOS app. |
|
|
| | | | |
| | --- | --- | |
| | File | `Spleeter2StemsVocalMask-F1536-fp16.mlmodelc` (compiled ML Program, FP16, macOS 14+) | |
| | Upstream | `2stems.tar.gz`, SHA-256 `f3a90b39dd2874269e8b05a48a86745df897b848c61f3958efc80a39152bd692` | |
| | Conversion | The two U-Nets re-implemented in PyTorch (matching TensorFlow at a median 72 dB SNR), then `spleeter_torch.py convert --f-bins 1536` with coremltools and `coremlcompiler` | |
| | Input | `magnitude`, float32 (1, 2, 512, 1536): a stereo magnitude spectrogram | |
| | Output | `vocal_mask`, float32, same shape: the vocal ratio mask, `sv² / (sv² + sa² + 1e-6)` | |
|
|
| The weights are unchanged apart from the conversion to FP16. |
|
|
| ## Licence |
|
|
| MIT licence; see `LICENSE`. Copyright (c) 2019-present Deezer SA. |
|
|
| > R. Hennequin, A. Khlif, F. Voituret and M. Moussallam, "Spleeter: a fast and efficient music source separation tool with pre-trained models", Journal of Open Source Software 5(50), 2154, 2020. |
|
|
