UX-Net Based Speech Separation Scheme using Parameterized Hypercomplex Convolutions
Samarpreet Singh, Uajjawal Sharma, Vinal Patel · 2024
This paper presents a modified U-Net-based approach for speech separation that employs parameterized hy-percomplex convolutions. The proposed model is capable of handling both single and multi-microphone input signals. It is for signals in both the frequency and time domains, and then utilizes a separation module to estimate the masks of the speech sources. The UX block filters the input across different resolutions to approximate these masks. The overall number of model parameters is reduced by leveraging the properties of Clifford algebra. Kronecker products are utilized for the parameterization of the hypercomplex convolution layers. The parameterized hypercomplex networks are able to subsume the algebraic rules from the data, thus enabling their application across domains ranging from I-D to n-D. Consequently, the number of network parameters can be reduced to l/n using these PHC layers, where n is a hyperparameter indicating the hypercomplex domain.