Research On The Method Of Feature Extraction And Endpoint Detection In Speaker Recognition And The Realization Of DSP

This paper focuses on the method of feature extraction in the speakerrecognition system, mainly researching on the performance of the single featureand the combined feature. Besides, an algorithm of speech endpoint detectionbased on expanded spectral subtraction and the SAP soft decision is proposed toimprove the capability of endpoint detection for the lower SNR situation. On thebasis of the study mentioned above, a speaker real-time recognition systemrealized with Digital Signal Processor(DSP) is designed and complished. Nowthis system has been used in a phonetic electronic lock.As a kind of biometric identification technology, speaker recognition is torecognize people's identity from its voice, which contains physiological andbehavioral characteristics specific to each individual. It has developed quicklywithin this decade and has important application in military affairs, law and bank,etc. With the development of Biometric Identification Technology, speakerrecognition has caught many eyes for its unique superiority, thus it becomes awidely used technique in human's life and work.DSP which is a kind of real-time microprocessor in digital signalprocessoring area has been widely used in automation, military affairs, radar,medical treatment and home appliances, etc. Its advantages are real-time andspeedy. When DSP is applied in speech signal processing, it always becomes agood measure in speech recognition, speaker recognition and speechenhancement.It is obvious that the research on the speaker recognition system and therealization of DSP has important value in theory and application.The main work and content are sum-up as followThe dissertation consists of six chapters.In Chapter 1, the research contents and principal work of the dissertation areexplained. Then the development and present situation information of BiometricIdentification Technology, speaker recognition and DSP are briefly reviewed.In Chapter 2, the basic knowledge of speaker recognition, the method offeature extraction, speech enhancement and endpoint detection are introduced. Atthe same time, the basic information of the software realization of DSP issummarized. All the above contents are the base of later research.In Chapter 3, the performance of single feature, further feature extractionand combined feature are mainly researched. MFCC, LPCC and their dynamicfeature are disscussed, then large numbers of experiments were taken to comparethe performance of these features under different orders. According to the resultsof these experiments, MFCC+ΔMFCC and LPCC+MFCC+ΔMFCC are chosenin practice.In Chapter 4, the endpoint detection of the noisy speech signal for the lowerSNR situation is introduced. An algorithm that combines the expanded spectralsubtraction with the SAP soft decision is proposed based on the study of thetraditional methods. The algorithm uses a method of expanded spectralsubtraction based on the noise compensation structure ruled by an adaptivedecision, which can estimate the noise during speech presence. At the same time,a method of end point detection based on the SAP(Speech Absence Probability)soft decision is given, which improves the robustness and the precision of theendpoint detection. The experiments show that higher efficiency and gooddetection capability can be obtained while the SNR equal to -10dB, however thetraditional method based on energy can't achieve such good performance underthe same SNR.In Chapter 5, the speaker recognition system is completed and the programis realized by DSP Also the debugging results of the programs which adoptdifferent features, distortions and recognition rules are given. After the test inMatlab and DSP, the most efficient one has been selected to apply in the phoneticelectronic lock.In Chapter 6, general conclusion is set out together with the further researchdirection and future work.The main contributions of this dissertation are listed as followⅠ Collect and clean up a lot of literature about speech signal processingand speaker recognition. Study the basic knowledge of DSP.Ⅱ Research on the single feature and combined feature while studying thebasic knowledge. Then the results of some experiments are given aftercomparing these different features. According to the results, a lot of valuableconclusion and experience can be gotten, which set a good basis for the DSPrealization of the algorithm.Ⅲ Research on the endpoint detection of the noisy speech signal for thelower SNR situation, and analyse the performance of the traditional method.Through the comparison, an algorithm that combines the expanded spectralsubtraction with the SAP soft decision is proposed. Therefore, the algorithm canprovide better performance.Ⅳ From the point of view of combined feature, study the speakerrecognition system and carry out the algorithm in Matlab. Then the algorithm isrealized with DSP and applied in the phonetic electronic lock.And this paper has some heuristic value on the following three questions inspeaker recognition and the DSP realization of algorithm.Ⅰ How to improve the performance of the combined feature and find moreefficient parts of the features for the sake of reducing the order and calculation infeature extraction.Ⅱ How to find more efficient method to estimate the noise in speechenhancement, so that endpoint detection can become simpler.Ⅲ How to find a new way that is more efficient and reliable to identify thespeaker, and make the realization with DSP easier.