DEEP CLUSTERING AND CONVENTIONAL NETWORKS FOR MUSIC SEPARATION: STRONGER TOGETHER.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Deep clustering is a novel approach for audio source separation, adept at handling multiple sources of the same type.
• Its efficacy in challenging scenarios like music source separation remains largely unexplored.
• Conventional methods directly estimate source signals, optimizing for signal approximation.

Purpose of the Study:

• To evaluate deep clustering's performance in music source separation, specifically singing voice separation.
• To compare deep clustering against conventional networks in both matched and mismatched conditions.
• To investigate the potential benefits of combining deep clustering with conventional network architectures.

Main Methods:

• Deep clustering generates time-frequency bin embeddings for source separation through clustering.
• Conventional networks are trained end-to-end for direct source signal estimation.
• A hybrid network combining both approaches was developed, utilizing multi-task learning principles.

Main Results:

• Deep clustering outperformed conventional networks in singing voice separation tasks.
• The hybrid network significantly surpassed the performance of either individual component.
• Deep clustering demonstrated robust performance even in mismatched conditions.

Conclusions:

• Deep clustering offers a flexible and effective alternative for audio source separation, particularly for complex musical content.
• Hybrid architectures integrating deep clustering and conventional methods show promise for advancing audio separation technology.
• The complementary strengths of deep clustering and conventional networks can be synergistically leveraged for enhanced performance.