Robust Speech Instruction Recognition And Embedded Implementation For Cold Voice

In recent years,speech recognition technology has gradually entered into people's daily life,and the Intelligent Home system which has joined the speech recognition technology makes people's life more comfortable and convenient.However,in the practical application scene,the accuracy of speech recognition is affected by many factors,such as environmental noise and the speaker's own tone(because of colds,vocal cord inflammation,pharyngitis,etc),which result in weak robustness of speech recognition system.Researches have paid extensive attention to the environmental noise factors and taken various measures to reduce the impact of environmental noise.However,there is still a lack of research on the tone of speakers themselves.Therefore,this paper studies the speaker's speech recognition under cold conditions,and focuses on the speech recognition system to ensure the normal speech recognition rate while making the cold speech recognition rate as close as possible to the normal speech recognition rate,in order to improving the performance of the speech recognition system.The main research work and innovation of this thesis are stated as follows:(1)A cold voice database was established.According to the Intelligent Home Control Voice Command,the "Pathological Abnormal and Normal Speech Data" was established.The database has been included in the Chinese Linguistic Data Consortium.(2)Differential analysis and the feature parameter processing are performed on the features of cold and normal voice.Using the speech of the subjects before and after colds,we analyzed the extracted feature parameters statistically.The results showed that the target audio rate,Resonance Peak and Mel Cepstrum coefficients of the subjects before and after colds have obviously differences.According to the difference of the feature parameters,this paper proposed a time normalization method based on the feature space trace,which uses the mean value of the speech signal in the segment to represent the speech feature.The experimental results show that compared with the Mel Cepstrum coefficients feature,the features processed by this method can effectively reduce the feature differences between normal and cold voice.(3)In view of the mismatch between cold voice and normal voice template,two kind of speech recognition schemes that are robust to cold voice are proposed.The first is a speech recognition scheme based on decision fusion,which uses SVM classifier to discriminate the normal voice and cold voice of input speech.If there is a little difference between cold voice and normal voice,the recognition result is obtained by the decision fusion method,otherwise,the recognition result is obtained by the corresponding speech template according to the classification result.This scheme makes up for the error caused by the speech discrimination process,and improves the recognition rate of cold voice in speech recognition system.The second is a speech recognition scheme based on time normalization of feature space trace,which uses time normalization to deal with cold voice feature parameters.This scheme can effectively improve the recognition rate and real-time performance of cold voice.(4)The embedded speech recognition system is realized.The system includes speech recognition algorithm module,a man-machine interface module and an online learning module.Finally,the system is tested under real-world conditions,and the system recognition rate is about 77.52%.