关键词: 语音分离, 调制机制, 孪生网络, 协同注意力机制, SOM

Abstract: To address the problem of overlapping time-frequency points among different speakers that leads to poor separation performance in single-channel speech separation methods based on spectrogram features, a single-channel speech separation model based on the Auditory Modulation Siamese Network is proposed. First, the modulation signal is computed through frequency band division and envelope detection, and the modulation amplitude spectrum is extracted using Fourier transform. Then, the mapping relationship between the modulation amplitude spectrum features and speech segments is obtained using a mutation point detection and matching method to achieve effective segmentation of speech segments. Subsequently, a Siamese Network based on the co-attention mechanism is designed to extract discriminative features of speech segments from different speakers. Afterward, a Self-Organizing Map Network (N-SOM) based on the domain influence mechanism is proposed, which performs feature clustering without pre-specifying the number of speakers by defining a dynamic neighborhood range to obtain mask matrices for different speakers. Finally, to avoid artifacts in the reconstructed signals in the modulation domain, a time-domain filter is designed to convert modulation-domain masks into time-domain masks and reconstruct speech signals by combining phase information. Experimental results demonstrate that the proposed model outperforms the Double-Density Dual-Tree Complex Wavelet Transform (DDDTCWT) methods in terms of Perceptual Evaluation of Speech Quality (PESQ), Signal-to-Distortion Ratio improvement (SDRi), and Scale-Invariant Signal-to-Distortion Ratio improvement (SI-SDRi) on both the WSJ0-2mix and WSJ0-3mix datasets.

Key words: speech separation, modulation mechanism, siamese network, co-attention mechanism, self-origanizing maps