Improving Speaker Recognition by Time-Frequency Domain Feature Enhanced Method

Many existing speaker recognition algorithms have the problem that single-domain feature extraction cannot represent the speech characteristics well, and this problem will affect the accuracy of speaker recognition. To solve this problem, we propose a time-frequency domain feature enhanced deep speaker (TFDS). The proposed algorithm can combine time domain and frequency domain, enhance the traditional MFCC feature extraction, and make up for the shortcomings of other algorithms that only extract features in a single domain. The deep speaker network architecture includes ResCNN, GRU, time averaging layer, style transformation layer, length normalization layer, and the loss is triple loss. Representation of experimental results performed on the librisspeech dataset results show that TFDS has higher accuracy and lower Equal Error Rate than deep speaker, and the time-frequency domain feature enhanced method can also be combined with other networks to improve the accuracy of speaker recognition.