| --- |
| license: mit |
| tags: |
| - beat-tracking |
| - downbeat-tracking |
| - core-ai |
| - aimodel |
| - apple-silicon |
| - macos |
| library_name: swift-music-transcriber |
| --- |
| |
| # Beat This! for Core AI (`.aimodel`) |
|
|
| Apple Core AI conversion of **Beat This!**, the beat and downbeat tracker by Francesco Foscarin, |
| Jan Schlüter and Gerhard Widmer (CPJKU) — *Beat This! Accurate Beat Tracking Without DBN |
| Postprocessing*, ISMIR 2024 — checkpoint `final0` from |
| [github.com/CPJKU/beat_this](https://github.com/CPJKU/beat_this) (MIT, 20M parameters). |
|
|
| `scribe-beat-this-float32.aimodel` (79 MB) takes a log-mel spectrogram `[1, T, 128]` at 50 fps |
| (22.05 kHz audio) and returns framewise beat and downbeat logits. The front end, the 1500-frame |
| chunking with 6-frame borders, the minimal peak picker and the tempo/metre fit live in |
| [swift-music-transcriber](https://github.com/arraypress/swift-music-transcriber), which uses this |
| as the beat tracker for its MIDI grids — the same tracker MuScriptor's own pipeline uses. |
|
|
| ## Faithfulness |
|
|
| One method (`Attention.forward`) was re-authored for the Core AI compiler, which rejected the |
| original einops reshape on a dynamic sequence length; RoPE became a constant matrix. The export |
| script asserts the re-authored module matches the original before exporting (max |diff| below |
| 1e-3 on random input, 0 in practice). End to end, beats land within one frame of upstream on the |
| reference clips; resampling (julius versus soxr) is the only step not shared with upstream. |
|
|
| ## Use |
|
|
| ```sh |
| hf download arraypress/scribe-beat-this --local-dir models |
| scribe model install models/scribe-beat-this-float32.aimodel # picked up automatically as the tracker |
| ``` |
|
|
| Reproduce with `uv run Tools/export_beat_this.py` in the library repo. macOS 27, Apple silicon. |
|
|
| ## Licence |
|
|
| MIT, as upstream. Please cite the ISMIR 2024 paper if you use it. |
|
|
