Hybrid Demucs with Self-Knowledge Distillation (PS-KD)

This repository introduces the Self-Knowledge Distillation with Progressive Refinement of Targets (PS-KD) approach to training Hybrid Demucs. The model training incorporates a curriculum learning approach. Additionally, it includes the codebase for the Cadenza Challenge 2024.

• Hybrid Demucs (V3) [Paper] [Repo]
• PS-KD [Paper] [Repo]

Music Source Separation Training Code

Training code for the mss models based on [here]

Results

• The results indicate the performance of the validation set for the cadenza challenge.

PS-KD

• The PS-KD method uses targets by creating soft targets, and we used an alpha of 0.8.

Curriculum-learning

• We took the SDR score and used the top 75% of the scoring datasets as the EASY dataset.

Augmentation

• We used a random augmentation method on pitch {-2, -1, 0, 1, 2} tempo {-20, -10, 0, 10, 20} with random augmentation. We also applied a random source mix, channel shuffle, and random audio effects (reverb, phaser, distortion) with a 0.05% probability. Soundstretch was used to augment pitch and tempo, and other techniques are available in code.