Research And Implementation Of Deep Belief Networks Based Speaker Recognition

With the rapid development of Internet Technology, speech signal processing technology has been applied in more fields. Speaker recognition technology is one of the important branches and the hotspot in the field of current speech signal processing technology. Speaker recognition is a technology about speaker identity authentication, which including speech recording, preprocessing, feature extraction, model training and pattern matching. It determines the speaker identity through model training and pattern matching of the speech features from the speech after pretreatment. In fact, feature extraction is the foundation of other steps and one of the most important part of the speaker recognition. Therefore, this thesis analyzes the advantages and disadvantages of feature extraction methods and puts forward an improved feature extraction algorithm of deep belief networks by improving parameter setting. In addition, there are many factors that can influence the feature extraction, in which noise is the most direct factor, so that speech filtering must be conducted before the feature extraction. Therefore, the advantages and disadvantages of speech filtering methods are also analyzed and a multilayer adaptive morphological filtering algorithm is proposed in this thesis. After that, a prototype system of speaker recognition is designed and implemented. The main contributions are listed as follows:(1)The characteristics and difficulties of speech filter algorithms and feature extraction algorithms are summarized, and the advantages and disadvantages of the current commonly used speech filter algorithms and feature extraction algorithms are analyzed.(2)Most of the existing filtering algorithm lost more clean speech signal during speech filtering, which degraded the speech quality. The morphological filtering algorithm can not only filter the noise in speech effectively, but also reduce the loss of clean speech. However, There are problems such as a fixed structure,pre-set structural elements and bias correction coefficient in existing morphological filtering algorithm, to improve these deficiencies, an adaptive filtering algorithm with a multi-layer structure is proposed. The proposed algorithm can choose and use different structure elements flexibly when facing the complicated changing interfering signal. In addition, aiming at the bias causing by the opening operation and closing operation, it can also weaken the negative impact that the bias correction coefficient vector. The simulation results show that the proposed algorithm improves the performance of morphological filtering, and it has the characteristics of simple design and strong practicability.(3)The processing of initial feature in existing feature extraction methods are simple combination, differential, selection or weighting of features, which cannot get high accuracy. To solve this problem, the deep belief networks is used for feature extraction, and the parameter setting is also improved to exploit theintrinsic link between the data fully so that the extracted features have a stronger characterizing ability. The simulation results show that the features can reduce the error rate of speaker recognition.(4)The demand for speaker recognition system is analyzed, and the partition of function, which is based on the operation process of speaker recognition, is implemented, and the speaker recognition prototype system framework is designed. In addition, this thesis designs and implements the core module of the system framework. Lastly, the system is tested in function-level and system-level, and the tested results are analyzed.