Research And Design Of Embedded Speaker-independent Isolated Words Speech Recognition System Based On DSP

Speech Recognition is the best way of the personal-machinecommunication. By the development of computer technology, it has beenimplemented into product in many areas such as control, communication andconsumption. Not only is speech recognition a theoretical problem, butalso an engineering problem. It integrates theoretic achievements of manydisciplines, for example, acoustics, phonetics, linguistics, physiology,digital processing, information engineering, computer science, patternrecognition and artificial intelligence. How to make the speechrecognition put to practical uses is the hotspots of current research ofspeech recognition technology.This paper mainly focus on speaker independent isolated word speechrecognition algorithm on embedded platform. The paper gives how toimplement speech recognition system on embedded platform which haslimited resource and low process capability, as to increase the system'sstability, decrease the system's response time and improve the system'sresource usage.Fundamentals of speech recognition and its algorithm have been studiedin this paper. We compare the difference of the speaker independentisolated word speech recognition algorithm, and select some effectiveapproaches for our system. Then we research on how to realize our speakerindependent isolated word recognition algorithm on DSP platform. Thealgorithm was realized and emulated by MATLAB in computer finally.The procedure of speech recognition mainly comprise pre-processingof speech signals, end-point detection, feature extraction, theconstruction process of speech corpus, pattern matching. In this speechrecognition system, the feature extraction algorithm is linear predictionanalysis (LPC), the pattern matching algorithm is dynamic time warping(DTW) and the construction process of speech corpus by clustering. Start and end point detecting is very important in isolated speech recognitionsystem. In this paper, we put forward a changeable show-time energy methodon the base of double restricted detection. The algorithm has beencertified to be efficient by MATLAB emulating and has a good precisionin this speech recognition system.