Enhancing the Performance of Single-Channel Blind Source Separation by Using ConvTransFormer

Abstract

In the specialized field of audio signal processing, this study introduces a pioneering ConvTransFormer architecture aimed at enhancing the performance of single-channel blind source separation (SCBSS). This innovative architecture ingeniously combines the strengths of a multiple simple-weak attention mechanism with the triple-gating feature of a Gated Attention Unit (GAU) within the ConvTransFormer. This combination allows for a more focused and effective targeting of specific segments within the input sequence. The efficacy of this ConvTransFormer architecture is rigorously evaluated using the WSJ0-2mix dataset, a standard benchmark in the field. The results of this evaluation are significant, demonstrating substantial improvements in key performance metrics. Notably, there is an increase in the Signal-to-Interference (SI)-Signal-to-Noise Ratio improvement (SNRi) by 16.5 and in the Signal-to-Distortion Ratio improvement (SDRi)-Signal-to-Interference (SDRi) by 16.8. These improvements are crucial indicators of the quality of source separation in SCBSS. The findings of this research are groundbreaking, indicating that the proposed ConvTransFormer architecture surpasses existing methods in both SI-SNRi and SDRi performance metrics. This advancement marks a significant step forward in the field of SCBSS, offering new avenues for more effective and precise audio signal processing, especially in scenarios where isolating individual sound sources from a single- channel input is essential.