| # Unpacked ASO example | |
| This is one 3.072-second training crop with 11 isolated notes at 16 kHz. | |
| `mixture.wav` is the input, `note-*.wav` are its aligned source waveforms, and | |
| `notes.json` contains pitches and sample-based note boundaries. The sources | |
| sum to the mixture within the dataset floating-point tolerance. | |
| A negative onset means a note was already playing when the crop began. Symbolic | |
| offsets come from the score; acoustic offsets include the measured release. | |
| All WAV files preserve the original float32 samples. | |