Research on Speaker Speech Segmentation and Clustering Based on Mixed Features in Complex Noise

As the front-end technology of speech signal processing, speech segmentation and clustering seriously affect the effect of speech recognition and speaker recognition.When audio data contains multiple speakers and complex background noise, the accuracy of target speech segmentation and extraction will decline sharply.To solve the problem, an auto-encoder based on U-Net is designed to reduce the dimension of speech mixed features, which can effectively improve the poor discrimination and robustness of speech acoustic features under complex conditions.A full attention module based on Res2Net is proposed for SKA-TDNN network, which significantly enhances the ability of the network to extract the speech features of shortduration speakers.According to the characteristics of speaker voice embedding features, an improved k-means speaker clustering algorithm based on peak statistics is proposed, which makes up for the defect of traditional k-means clustering algorithm that the results are easily affected by the selection of initial clustering centers, and further improves the accuracy of clustering results.At the same time, under the complex background noise of low SNR, the segmentation and clustering algorithm based on mixed features is designed.The experimental results show that the recall rate, accuracy rate and comprehensive performance are significantly higher than the traditional algorithm, which has stronger robustness and antinoise ability.