Speaker Recognition Based On DTW And Its Realization On DSP

Speaker recognition technology is an authentication technology developed with the IT,which involves multi-disciplinary research area, advances in these areas and promote thedevelopment of this technology. With the arrival of the era of intelligent, people hopecomputer and other intelligent devices can be more convenient，reliable communication.However, most of the small portable intelligent devices need to enter your password easyto forget and difficult to use, The speaker recognition just to make up for shortcomings inthis regard. In recent years, the speaker recognition have been achieved high accuracy intheory and laboratory conditions, and began to enter the practical application stage.Therefore, the embedded DSP system based speaker recognition has become a hot topic ofthe present study.This article introduces the background of the speaker recognition, significance, andits fundamental concepts and classification，including the principle of speech production,the analysis method of the speech, commonly used features in speaker recognitionparameters and their extraction methods, as well as commonly used in speaker recognitionmethod, then presented the hardware platform for system development, detailed focusingon pretreatment, endpoint detection, feature extraction and pattern matching in process ofthe speaker recognition. Finally after transplant the designed algorithm of system on a DSPplatform, through statistics used operation time and storage space in the various parts ofthe algorithm to analysis the performance of the whole system.This paper implements a6speaker identification system on DSK5402board of TICo，According to the requirements of the system, after compare and analysis thecharacteristic parameters and identification algorithms of commonly used speakerrecognition. Features the system used include Combination parameter consist of MelFrequency Cepstrum Coefficient and Speech frame energy，Matching algorithm of theimproved DTW. Then simulate the algorithms that used by the system with MATLAB, in order to determine the reliability of its performance, and introduce linear predictive speechsynthesis methods to achieve the prompt for identification statement and output therecognition result, to save more system memory, faster and more convenient to use. Finally,adopts such means as the conversion of fixed-point and floating-point data, change thelength of data access and use of inline functions to optimize the software systems, andsuccessful implementation of the bootstrap run. Experiments show that the system achievessatisfactory results on speaker identification test with appointed speech content.The system easy to use, has a faster recognition rate and is basically meet the designrequirements, however, there are some areas for improvement. For example, the system istested in the laboratory environment, and did not consider the problem of noise immunity,in further studies, we should put the effect of noise into account. In addition, our systemonly applies to the text-related speaker recognition, therefore the use scope is relativelynarrow, the next step should be designed from the perspective of text-independent, so thatthe system can be applied to more areas.