Speaker Recognition on Mobile Phone: Using Wavelet, Cepstral Coefficients and Probabilisitc Neural Network

With the used of speech as a biometric in remote login or register, speaker recognition is widely applied on the smart mobile phone. However, the accuracy of speaker recognition usually drop off rapidly for the rough speech signal transmitted by the phone channel. This paper propose a new robust speaker recognition model for mobile phone context. The cepstral coefficients and its delta and delta-delta coefficients are extracted from the signal as the feature vector. Moreover, the discrete wavelet transform and threshold de-noising process are employed to enhance the performance of the feature in the noisy environment. The extracted feature vectors are used as input to the probabilistic neural network, which is a faster classifier compared with the back-propagation neural networks. To simulate the speech transmitted over the mobile phone channel, the TIMIT corpus is downsampled to 8KHz, and 20dB Gaussian white noise is added into the corpus. The experimental results show that the proposed model obtain high accuracy and can increase the true positive rate for a given false positive rate in a noisy environment. Therefore, the proposed model is suitable for mobile phone context.