Application of Voiceprint Recognition Based on Improved ECAPA-TDNN

In speech signals, the characteristics of the same type of voiceprint in time usually have similar trends, and the trends of different voiceprints in time are quite different, which makes the time context information have a great influence on the results of speech recognition. Although the ECAPA-TDNN model can obtain time context information through dilated convolution to a certain extent, the time context information obtained by the model is not sufficient. Based on the ECAPA-TDNN model, this paper adds a bidirectional long-term and short-term memory network (Bi-directional LSTM, BLSTM) that can fully mine the time context, thus proposing the ECAPA-TDNN-LSTM model. The proposed model can not only emphasize channel attention, but also have the characteristics of information dissemination and aggregation, and can obtain rich time context information. Experimental results show that the recognition accuracy of the proposed algorithm reaches 89.17%, which improves the recognition accuracy of voiceprint.