Beat This! final0 — ONNX for Key & Tempo
This repository distributes the FP32 ONNX export of the official Beat This! final0 model used by Key & Tempo.
This is a conversion of the upstream pretrained model, not a newly trained model and not an official CPJKU release. It predicts beat and downbeat logits; BPM, time signature and MIDI tempo maps are derived by the application.
Files
| File | Purpose |
|---|---|
beat-this-final0.onnx |
FP32, opset 17 network, 82,098,846 bytes |
frontend.json |
Sample rate, FFT settings, window and mel filter coefficients |
manifest.json |
Source commit, size, SHA-256 and export parity results |
LICENSE |
Original upstream MIT license and copyright notice |
Input and output
The network consumes a log-mel spectrogram, not raw audio. Input spectrogram has shape [1, frames, 128]; outputs beat and downbeat have shape [1, frames]. The frontend uses 22,050 Hz mono audio, a 1,024-point FFT and a hop length of 441 samples. Use the accompanying frontend constants and the application preprocessing, chunking and postprocessing implementation for compatible results.
See the export script and browser inference code. Hugging Face hosts the files only; the application performs inference in the user's browser and does not upload audio to Hugging Face.
Provenance and validation
- Original project: https://github.com/CPJKU/beat_this
- Source commit:
b95c8ab0c58c2d9fcfd40508ae8dffbc05ac4f5c - Pretrained checkpoint:
final0 - Model SHA-256:
e5e37b7d1802895e42559c5a1ced7619b1601d257a5c8e167c9d45de37c1b078 - Export parity: compared against native PyTorch at 63, 128, 1264 and 1500 input frames; results are recorded in
manifest.json.
The checks establish conversion consistency, not a new accuracy benchmark. Beat tracking may fail, and downstream tempo estimates can reflect half/double tempo, missed beats or timing jitter. The model does not estimate musical key; Key & Tempo uses a separate S-KEY model for that feature.
License and attribution
Copyright (c) 2024 Institute of Computational Perception, JKU Linz, Austria.
The upstream README states that its code and published model weights are MIT-licensed. The complete original license is included in LICENSE; retain it when redistributing this model. Training audio has separate rights and is not included here.
Citation
The upstream README supplies this citation:
@inproceedings{foscarin2024beatthis,
author = {Francesco Foscarin and Jan Schl{\"u}ter and Gerhard Widmer},
title = {Beat this! Accurate beat tracking without {DBN} postprocessing},
year = 2024,
month = nov,
booktitle = {Proceedings of the 25th International Society for Music Information Retrieval Conference (ISMIR)},
address = {San Francisco, CA, United States},
}