MAGDA Beat Tracker
ONNX export of Beat This! (CPJKU), packaged for the MAGDA DAW's media library (issue #2674).
What's in this repo
| File | Size | SHA-256 |
|---|---|---|
beat_this.onnx |
79.2 MB | c5c1466e08abdb03fdeb50668a06f244b787d564c212490482231a9cfbe9ccbd |
A transformer over log-mel frames. It takes a whole file at once and returns a beat and a downbeat logit per 20 ms frame; MAGDA reads a tempo off the spacing of the beats and a bar length off the spacing of the downbeats.
Input contract
The mel front end has to match the one the model was trained with, or the frames mean nothing to it:
| sample rate | 22050 Hz, mono |
| FFT / window | 1024, Hann |
| hop | 441 (50 frames a second) |
| mel bands | 128, Slaney scale, 30 Hz - 11 kHz, unnormalised triangles |
| spectrum | magnitude, divided by sqrt(window length) |
| compression | log1p(1000 * x) |
| tensor | input_spectrogram, [1, frames, 128] |
| outputs | beat, downbeat, both [1, frames] logits |
How MAGDA uses it
The media database's third BPM tier. A file's tempo is read from its filename first and its ACID chunk second; when neither says anything, this measures it. An answer is kept only when the beats it found are steady, because a wrong tempo seeds a clip and stretches it while a missing one leaves a field blank.
Without the model MAGDA falls back to an autocorrelation over the spectral flux envelope it computes anyway -- which is right far less often, and says so.
Provenance
- Model and weights: CPJKU/beat_this, MIT.
- ONNX conversion: mosynthkey/beat_this_cpp, MIT, converted from the upstream PyTorch checkpoint with opset 14.
License
MIT, as upstream. Note the upstream caveat: some of the material the model was trained on is copyrighted or under limited Creative Commons terms, and it is on the user to decide whether that bears on their use.