Research On Mixed Sound Processing And Spam Text Filtering In Speech Recognition

The traditional speech recognition system is basically based on a relatively pure speech environment,sound capture,simple filtering and denoising,and then directly input the sound signal into the speech recognition network to obtain the recognized text.Speech recognition is severely affected once there is noise in the environment or the presence of other people's voices.Existing methods such as signal denoising,parameter denoising and anti-noise recognition have certain limitations.If the automatic separation of interference,noise and speech can be achieved,the relatively pure speech is obtained before the recognition,the recognition problem in the noise environment can be completely solved.The rapid development of technologies such as sound source separation in recent years has made it possible to separate interference,noise and speech.At the same time,in the process of a human-machine dialogue,the robot simply searches for the "response text" by the "question text",and then turns the "response text" into a sound,and the robot answers the "question text" that is heard.Without any resolving ability,robots' random answers often occur.For these two issues,this thesis proposes a speech recognition system design method based on blind source separation and junk text filtering.Firstly,the two methods of PCA and ICA are used to realize the blind source separation of the mixed sound stream,and the influence of the blind source separation method on the recognition of the mixed sound stream is tested.It is concluded that the addition of blind source separation can improve the recognition effect of speech recognition system on mixed sound stream to some extent.At the same time,it is found that the success rate of ICA method to achieve blind source separation can reach 96.4%,which is higher than the 89.6%success rate of PCA method.Three text filtering implementation methods are proposed,namely rule-based text filtering,Bayes-based text filtering and SVM-based text filtering.Experiments show that text filtering greatly reduces the false response rate of the speech recognition system to the "junk text",and the recall rate of the three methods for "junk text" can reach 92.4%,89%and 98%respectively.Therefore,this thesis uses the ICA method to achieve sound source separation,and uses the SVM method to achieve text filtering.The sound source separation can separate the mixed sound stream heard by the robot into several pure sound streams,so that the speech recognition system can separately identify and process each pure sound stream,avoiding the low recognition rate of the mixed voice recognition of the traditional speech recognition system.Text filtering can judge whether the identified text corresponding to each pure sound stream is qualified,and filter out the "junk text" that is not worth answering,to avoid the problem of false response to a certain extent.The problem of misrecognition caused by text filtering can judge whether the identification text corresponding to each pure sound stream is qualified,and filter out the "junk text" that is not worth answering,to a certain extent.The workflow of speech recognition system based on blind source separation and text filtering is seven steps:sound monitoring,sound capture,sound filtering,sound source separation,speech recognition,text filtering,and text response.Experiments show that compared with the traditional speech recognition system(sound monitoring,sound capture,sound filtering,speech recognition,text response),the new speech recognition system has significantly improved the recognition rate of mixed sound streams,while at the same time,the response rate to "junk text" is significantly reduced.