A method of noisy Tibetan speakers verification based on SKA-TDNN

Abstract

In recent years, the research of speech enhancement application technology has important practical value. At the same time, speaker recognition is widely studied and used as a very valuable biometric recognition technology. However, the application of speech enhancement and speaker recognition technology in Tibetan is few. Due to the lack of speech enhancement back-end processing, the speaker recognition system may lead to low speech quality and low intelligibility. In this paper, the backend structure framework of SKA-TDNN network and Wave-U-Net model is constructed to realize noisy Tibetan speaker verification, and the model is improved and optimized. For speaker recognition, we use the model structure of SKA-TDNN in combination with multiscale SKA (msSKA) to better model utterances with different durations. In the back-end processing of speech enhancement, we use WAVE-U-NET structure and introduce further improved optimization architecture. Experimental results show that the improved SKA- TDNN model in speaker verification than traditional investigate VGG model was reduced by 5.225%, than ECAPA -TDNN model was reduced by 1.5%, and got close to 0.94 STOI in speech enhancement.