aaatmy/skey-onnx

S-KEY — ONNX for Key & Tempo

FP32 ONNX conversion of the pretrained Deezer S-KEY model, used by Key & Tempo to estimate a single global major or minor key from audio.

This is a community conversion of the upstream checkpoint, not a newly trained model or an official Deezer release. Hugging Face hosts the files; the application runs inference locally and does not send audio to Hugging Face.

Files

File	Purpose
skey.onnx	FP32, opset 17 model, 324,536 bytes
manifest.json	Hash, source commit, sample rate, ordered labels and parity results
skey-keys.json	Application mapping for all 24 classes
LICENSE	Original S-KEY copyright and MIT license text
ConvNeXt-LICENSE, nnAudio-LICENSE	Notices for architecture and preprocessing components

Input and output

Input audio is float32 mono PCM at 22,050 Hz, shape [1, samples]. Output scores has shape [1, 24]. Peak normalization, VQT/dB preprocessing, frequency cropping and ChromaNet are included in the graph. Scores already include the upstream softmax; do not apply softmax again.

Use the exact label order in manifest.json: major classes start at A Major, minor classes start at B minor. G Major is index 10; A minor is index 22. Key & Tempo requires at least 3 seconds and rejects silent input before inference. This minimum is an application policy.

Browser inference uses ONNX Runtime Web WASM CPU. See the export script, browser implementation and technical notes.

Provenance and validation

• Source commit: 918b83d273568d5041569bb8068843d19a335726
• Upstream checkpoint: skey/models/skey.pt
• Checkpoint SHA-256: 78dfd0ad4fa9434bf7cec70a25934b7c575bda9c80e994700140770ad3a5ead4
• ONNX SHA-256: 50bac66d5b02868afad07a0b159f08495cc3f5d8492654587e828ed4686412b1

The export rewrites affine-free LayerNorm using mean/variance operations and width-one adaptive average pooling using a time-axis mean for dynamic audio length. Checks compare the original model, rewritten copy and ONNX Runtime. The manifest records native/ONNX comparisons for approximately 3-, 25- and 75-second fixtures, with maximum score difference below 0.000004 on these examples.

These checks establish conversion consistency, not musical-key accuracy. The model estimates one global key, does not locate modulations or recognize chords, and its softmax scores are not calibrated confidence. Full-track inference can use substantial memory; the application's 20-minute limit does not guarantee every device can process that duration.

License and attribution

Copyright (c) 2019-present, Deezer SA.

The upstream README explicitly describes the S-KEY code as MIT-licensed. That paragraph does not separately state a license for pretrained weights; this model card does not assign them a new license. The original license is preserved in LICENSE. Retain upstream copyright and license notices when redistributing covered material.

S-KEY includes ConvNeXt-derived code and nnAudio preprocessing. Their bundled license texts are also retained. Sources: ConvNeXt, nnAudio.

Citation

The upstream README asks research users to cite:

@INPROCEEDINGS{kongskey2025,
  author={Kong, Yuexuan and Meseguer-Brocal, Gabriel and Lostanlen, Vincent and Lagrange, Mathieu and Hennequin, Romain},
  booktitle={ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  title={S-KEY: Self-supervised Learning of Major and Minor Keys from Audio},
  year={2025},
  pages={1-5},
  doi={10.1109/ICASSP49660.2025.10890222}}

Related references: A ConvNet for the 2020s, CVPR 2022; nnAudio, IEEE Access 2020.