On-the-fly pitch and tempo shift augmentation improves separation model generalisation across musical keys and tempos

MUSDB-HQ contains 150 songs drawn from specific genres, keys, and tempos. Demucs’s RepitchedWrapper applies random pitch shifts (±2 semitones) and tempo changes (±12%) to each stem during training, using soundstretch (a phase-vocoder-based time-pitch tool). Crucially, the same pitch/tempo change is applied consistently across all stems in a track so the reconstructed mix remains coherent. Vocals get special treatment via the -speech flag in soundstretch, which uses a speech-optimised algorithm that preserves formants better than musical mode.

Examples

From demucs/repitch.py:

def repitch(wav, pitch, tempo, voice=False, ...):
    # pitch in semitones, tempo as % delta
    # voice=True uses soundstretch -speech flag

Assessment

Why is it important to apply the same pitch shift to all stems simultaneously? What would happen if drum and bass stems were shifted by different amounts?