Application Research Of Spectrogram On Pronunciation Recognition Of Chinese Characters And Speaker Recognition

Speech recognition is a kind of easily acceptable biometric recognition technology,it has developed rapidly in recent years,and been widely used in security,voiceprint,voice locks,forensic identification.Chinese pronunciation recognition and speaker recognition are two main directions of speech recognition research,both of which are to process speech signals and extract features for recognition.Among them,the pronunciation recognition of Chinese characters is the extraction of the pronunciation characteristics of Chinese characters from speech signals to judge which Chinese character is pronounced;Speaker recognition is to extract speech signals,which reflects physiological characteristics of speaker's articulation system,and to automatically identify the speaker.At present,although speech recognition has been mature,it needs further improvement in recognition accuracy,recognizable sample size and recognition speed.The existing speech recognition technology usually takes MFCC,LPCC and LPMCC as the characteristic parameters.The recognition algorithm uses the hidden Markov model,vector quantization and dynamic time warping.Spectrogram is a two-dimensional image that shows the change of voice spectrum over time.It can not only highlight the overall time-frequency characteristics of speaker speech,the voice signal can also be visualized and the time and frequency characteristics of the speech be intuitively displayed.It also contains information that can be used for Chinese pronunciation recognition and speaker recognition.The experiments were completed on MATLAB2010 a software and the speech samples used in the experiment were recorded by the students in the quiet environment of the laboratory.In the study of the speech recognition of the isolated characters of Chinese characters,1605 Chinese characters in the commonly used Chinese character library are selected,and 920 pronunciations are taken as recognition objects.10 samples were obtained for each pronunciation so as to a total of 9200 samples are collected to train the convolution neural network.The experimental results show that the recognition rate of all samples reaches 99.32%,which is superior to other Chinese character pronunciation recognition methods.After the speech recognition of Chinese isolated words is achieved through the combination of spectrogram and deep convolution neural network,speaker recognition is carried out for 30 students in the laboratory.First,we use a complete syllable as a time unit to make a spectrogram for a speaker's voice,then we use the image processing technology to superimpose all the spectrogram obtained by each speaker to get the statistical characteristics of the speaker's pronunciation characteristics.Finally,we use superimposed spectrogram to train the convolution neural networks to judge which speaker is talking.The test results show that the recognition rate of all speakers reaches 98.83%,which is superior to other speaker recognition methods.The research in this paper has reference value for spectrogram and convolution neural network in speech recognition.