Research On Speaker Recognition Based On Gender Distinction

Speech signals contain both semantic information and personal information; itcould be extracted identity information from the signal, such as gender, age, nativeplace. As a kind of biological certification technology, speaker recognition adoptsspeech coefficients extracted from speech signal and makes an important applicationvalue and broad application prospect in information retrieval, public security crimes,speech identity authentication, phone banks and other areas. This paper study up on dataacquisition, feature extraction, classification identification and made the followinginnovative achievements.1. Set up a Chinese dialect speech databaseReferring to design standards of speech database in the world, we improved therecording channels, types of dialects, distribution of speaker’s age， gender andeventually set up a Chinese dialects speech database which contains seven kinds ofdialects (Min, Yue, Wu, Xiang, Gan, Northern, and Hakka) and mandarin. Includingbroadband voice (mic) and narrowband speech (mobile, fixed telephone); totalize106hours of speech data.2. Put forward a new gender identification method by using the codebook modelIt is the first time that introduces the semi-supervised vector quantization. Thespeech data is quantized by using semi-supervised learning principle and gendercodebook models of male and female with supervision information is formed. Becausethe method fully considered the probability distribution of the phonetic features statespace, and optimize the generate method of codebook, the new method greatly improveaccuracy of codebook model, solve the deficiency of low precision of codebookeffectively of traditional VQ model, and improved the system identification effectly.The test experiments under the noise voice and pure speech environment show that,compared with the traditional VQ algorithms, the new method with gender supervisioninformation are effectively improved in accuracy, stability and robust of the system.3. Put forward speaker recognition method based on mixed support vectormachinesSupport vector machine obeys structural risk minimization, strong ability todistinguish between classes. Output results reflect the differences between differentsamples, suitable for dealing with the classification problem of continuous input vectors.For this reason, we put forward the mixed SVM model system for speaker recognition.On the basis of segmentation and clustering large number of data samples, the method construct a SVM training for every type of speech, finally, make decision andclassification by comprehensive all the SVM output results. The new idea is a good wayto deal with problems of the system running time too long, low identify efficiency,which brought by the increase in the speaker numbers and large speech database, andeffectively improve the classification and decision-making ability of speakerrecognition system.4. Present a new method on stratified speaker identificationAt present, it is difficult to achieve speaker recognition of real-time applicationunder large speech data. As the expansion of speech database, the recognition systemcan not meet the needs of real-time identification in running time, memory requirementsand accuracy based on the existing technology. This paper firstly discusses differentperformance in the recognition system with several features such as MFCC, SDC. Thenaccording to the classification search method, narrows the numbers and ranger ofspeech data by using dialects identification, gender identification technology in speakerrecognition system. Finally, determine each speaker’s identity. We are seeking toestablish an optimal speaker recognition system model.