Research Of Speaker Feature Extraction And Recognition Algorithm

Due to its special merits of flexibility, economy and accuracy, speaker recognition was widely applied in the field of public security, judicature, voice-controlled cryptography, criminal verification, patient verification and military affairs. The research hotspots of speaker recognition are in two aspects: one is how to extract the effective speech parameter from large-scale speech data; the other is the design of effective classifer.Based on the problems existed in SVM and KFD, we proposed some novel and improved methods. The research work of this paper is as follows.1. Based on the deeply research of optimal dimension reduction and classify in PCA, we proposed the PCS-PCA classifier and TES-PCA classifier to solve the problem of high dimension and time-consuming model training existed in speaker recognition. This approach could consumedly reduce the computation complexity of later stage.2. Aimming at the problem of SVM that it could not process dynamic time sequence directly, add original information fisher score (AOI-Fisher score) was proposed to achieve sequence features. For the sake of reducing the compute complexity of AOI-Fisher score, PCA was utilized to reduce the dimensions of MFCC, and possible target speakers were selected synchronously. So the range of registered speaker was shortened and the number of input vectors was reduced. SVM could classify the whole sequence by this method; meanwhile the computation complexity was reduced.3. A hierarchical speaker verification method based on PCA classifier and kernel fisher discriminant was proposed. Using the simpleness and fast classification of PCA, the dimensions of registered speakers feature vectors was reduced, and possible target speakers were fast selected simultaneously. Then, KFD was used to make final decision. This method overcomed the limitation of KFD in which the compute complexity was enhanced along with the augment of speech data. Compared with conventional SVM and KFD, the experiment results showed that our proposed hierarchical classifier has superior performance, and this method is in the ascendant of recognition rate and model time.