thebluescreen/beat-this-onnx

Beat This! (ONNX)

Beat and downbeat tracking: framewise beat and downbeat logits at 50 frames per second for any piece of music, with no fixed tempo range or meter. A conversion to ONNX of the final0 checkpoint of Beat This! (Foscarin, Schlüter, Widmer, ISMIR 2024), with the log-mel front end fused into the graph.

• Input waveform: [1, N] float32, mono, 22.05 kHz, values in [-1, 1], where N = (frames - 1) * 441 + 1024. It is the audio around a run of frames spectrogram frames: the frame at index i is centred on sample i * 441, so the run [a, b) needs samples a * 441 - 512 to (b - 1) * 441 + 512. At the very start and end of a piece, extend by reflection (the centred STFT's reflect padding); everywhere else use the real neighbouring audio.
• Input pad_frames: [2] int64, the number of all-zero spectrogram frames to put before and after the run: 6 at the start and end of a piece, as the reference does, otherwise 0.
• Outputs beat and downbeat: [1, T] float32 logits, T = pad_frames[0] + frames + pad_frames[1]. beat already includes the downbeat term, as in the PyTorch model.
• Chunking, as in the reference: 1500-frame chunks that overlap by 6 frames on each side, the last chunk moved left to end at the end of the piece; drop 6 frames of each chunk's output at both ends and keep the first prediction where chunks overlap.
• A beat is a peak of the beat logits over a window of 7 frames with a logit above 0; peaks one frame apart are merged; a downbeat is moved to the nearest beat. Matches the PyTorch model to within 1e-4 on the conversion fixtures.

Converted by the Silly MIDI Tools scripts in scripts/beat-this/.

See NOTICE.md for the attribution and the changes made, and LICENSE for the licence text.