The Research And Design Of An Embedded Isolated-word Speech Recognition System

With the development of science and technology,speech recognition has been widely studied.This technology has been applied to embedded devices by more and more researchers,which has promoted the rapid development of speech control.At present,embedded speech recognition systems generate the matching model by training a large number of voice commands to complete speaker-independent speech recognition.However,systems specific for the household can only recognize Mandarin or specific dialect,which brings great inconvenience to users with heavy accent or poor Mandarin.Therefore,this thesis proposes a lightweight embedded speech recognition system.Since control commands can be customized,the system can be used not only for Mandarin but also for any dialects.Besides,it does not need to collect large amounts of voice data for training,which makes it to be a cost-saving device with wide applicability and high flexibility.This thesis first studies the preprocessing and endpoint detection of speech signals.A dual-threshold decision method based on short-time energy and zero-crossing rate is proposed and then optimized according to the usage scenario.After that,this thesis proposes a speech feature extraction method based on the combination of MFCC parameters and differential parameters.Because the embedded platform lack of resources and the calculation efficiency is low,we simplify FFT operation and propose a method without frame shift,which greatly reduces the computation of characteristic parameters.Next,the DTW algorithm matching process is optimized by using the double-template and threshold control method.Finally,we transplant the algorithm into AZ3166,and make some improvements according to embedded programming and system framework.A series of tests have been conducted to measure the proposed system both from its hardware and software.Experimental results show that the rate of correct speech recognition of this system is 91% with real-time response coefficient being 0.55,which fully proves its high value for practice.