A Method Using Deep Belief Network For Speaker Recognition Based On Feature Subspace Quantization

The main purpose of speaker age recognition is to infer the speaker's age range.The research of speaker age recognition system is a major respect in the side of speech processing.The characteristic parameters used in the existing age recognition system,such as linear prediction coefficient(LPCC),Mel-Frequency Cepstrum Coefficient(MFCC)are generally based on the short-time smoothness of voice signals,lacking description of dynamic characteristics.The spectral recognition according to Gaussian mixture model(GMM)is widely used.However,the traditional GMM transformation deals with every feature vector independently,ignoring the continuity among frames.In addition,the GMM-based age recognition system is also affected with over-smoothing because of weighted average.This paper proposed a new speaker age recognition methods based on deep belief network for age recognition.The input speech' Mel-frequency cepstrum coefficients which is called Wavelet Packet Mel-Frequency Cepstrum Coefficient(WPMFC)was got by wavelet packet transform.The speaker age is divided into four age groups such as children,youths,adult and older,the Feature Subspace Quantization is token for each group,and Gaussian mixture models are trained for each gender.Testing speech recognition decision is realized by maximum likelihood criterion.The neighbor frames of the training speech were gathered to form super-frames to serve as the input data of network,and the corresponding age categories were used as the output data of the network.The network used for age recognition was got by training.The results of experimental prove that the performance of age recognition based on proposed algorithm is successful compared with traditional GMM and DBN method.