Multi-dimensional Speech Feature Parameter Visualization And Its Application In Speech Recognition Research

With the rapid development of big data and e-commerce,multi-dimensional data is growing explosively,which brings new challenges and opportunities for the development of multi-dimensional visualization technology.In recent years,multi-dimensional visualization technology has made great progress and has been widely used in various fields.Multi-dimensional visualization technology can also be applied in the field of speech,but the traditional multi-dimensional visualization method can not display multi-dimensional speech feature data visually,and it is difficult to meet people's demand for multi-dimensional speech feature parameters visualization.In order to study the problems in speaker recognition and speech recognition better,a new multi-dimensional visualization method is proposed in this thesis,which can display multi-dimensional speech feature parameters visually.The main contents of this thesis are as follows:(1)The multi-dimensional visualization technology is studied,and the problems in multi-dimensional visualization technology are expounded.The basic principles,advantages and disadvantages of the existing multi-dimensional visualization technology are introduced in detail.A new multi-dimensional visualization method is proposed,which is based on the visualization of three-dimensional data.The method uses the dimension-by-dimensional expansion method to display data hierarchically,which can obtain visualization results better.(2)Research on multi-dimensional speech feature visualization in speaker recognition.Firstly,this thesis introduces the main steps of multi-dimensional feature visualization of speech signals.Then,the experiment of speaker recognition is carried out to study the influence of endpoint detection on speaker recognition rate,and the reason why endpoint detection can improve the recognition rate is analyzed by multi-dimensional visualization method.Finally,the method of multi-dimensional visualization proves that the distribution of test samples and training samples should be consistent in speaker recognition.Otherwise,it will seriously affect the recognition performance.(3)Research on multi-dimensional speech feature visualization in speech recognition.This thesis introduces the basic principle of speech recognition technology.Through the establishment of a model for each single Chinese character,the speech recognition experiment is carried out to study the influence of the order on the recognition rate.At the same time,the multi-dimensional visualization is used to analyze the reason why the order affects the recognition rate.Finally,it studies the multi-dimensional visualization of Chinese phonemes which are easy to be misunderstood in speech recognition,and analyze the cause of false recognition.The experimental results show that phonemes play an important role in speech recognition.By analyzing the phonemes visually and finding that the Chinese vowel parts are the same or similar,the distribution of the phonetic theory models is roughly the same.The vowel part contributes a lot to the theoretical model.When it contains the same or similar vowels,it is easy to misunderstand.