Optimizing Music Source Separation In Complex Audio Environments Through Progressive Self-Knowledge Distillation
This technical report presents our approach for The ICASSP 2024 SP Cadenza Grand Challenge (CADICASSP24), focusing on effective source separation. In the scenario addressed by this challenge, signals captured by hearing aid microphones are complicated by the intertwining of stereo signals, increasing their complexity. In such situations, where simple training losses like L1 loss can lead to significant errors and complicate model training, we introduce an effective fine-tuning method that softens the target using predictions from the previous epoch’s model. Our system improved the SDR score by 1.2 dB over the baseline. The source code is available online1.
Paper
Full text
Optimizing Music Source Separation In Complex Audio Environments Through Progressive Self-Knowledge Distillation
Semantic Scholar · Computer Science · 2024
Abstract
This technical report presents our approach for The ICASSP 2024 SP Cadenza Grand Challenge (CADICASSP24), focusing on effective source separation. In the scenario addressed by this challenge, signals captured by hearing aid microphones are complicated by the intertwining of stereo signals, increasing their complexity. In such situations, where simple training losses like L1 loss can lead to significant errors and complicate model training, we introduce an effective fine-tuning method that softens the target using predictions from the previous epoch’s model. Our system improved the SDR score by 1.2 dB over the baseline. The source code is available online1.