The article presents a new approach to speech separation using Single-Path Global Modulation (SPGM). Traditional dual-path models separate long sequences into overlapping chunks, modelling local and global features separately. However, it was found that the global modelling, which accounts for half of the model’s parameters, contributes minimally to performance. The proposed SPGM block replaces the global modelling with a more efficient approach, requiring only 2% of the model’s total parameters. This allows all transformer layers in the model to focus on local feature modelling. The SPGM model outperforms the traditional model in tests, achieving higher performance with fewer parameters.
Publication date: 25 Sep 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2309.12608
