Speaker identification using wavelet Shannon entropy and probabilistic neural network

摘要

Speaker identification is a technology widely used in security applications based on phone services. However, its performance is not very good because of the low quality speech transmitted over the telephone channel. This paper firstly proposes a new type of speech feature based on wavelet and Shannon entropy, and then combine the proposed feature with probabilistic neural network to present a new speaker identification model. The main advantage of our model is that it can take advantages of wavelet, probability neural network and Shannon entropy to obtain good performance on the condition that quality of speech is low. In our model, the speech is decomposed into 8 different subbands by discrete wavelet transform, and then 8 Shannon entropies are extracted from those subbands to form the feature vector. Finally, the extracted feature vector is used as inputs to a feed-ward neural network named probabilistic neural network(PNN). The TIMIT speech database is used to evaluate the proposed model. Compared with MFCC+GMM and ECD+GMM. The experimental results show that The proposed model obtained the best performance for low quality speech. Therefore, our new speaker identification model is suitable for speaker identification.