Study On Multi-feature Fusion Chinese Tone Recognition Algorithm Based On Machine Learning

Language is the most important,commonly used,and convenient way of communication for human beings.Chinese is a language with tones,which reflect the change in the height of human voice during pronunciation and play a role in the discrimination of meaning in people's daily communication.Tone recognition is a typical recognition problem of four or more classifications for the tonal change patterns of Chinese syllables.It has important applications in Chinese human-computer speech communication and language training systems for hearing-impaired patients.Therefore,tone recognition of speech signals is a meaningful research topic.The study of tone recognition is divided into two aspects:feature parameters and tone classifiers.In order to improve the accuracy,robustness and response speed of the tone recognition system,feature parameters should have the characteristics of high distinguishability,low computational complexity,and clear physical meaning,at the same time,tone classifiers should also have the advantages of strong classification ability,good anti-interference,and efficient operation.Study of this thesis starts from the above two aspects.First,the fundamental frequency parameter and seven typical feature parameter sets are introduced,and experiments prove that the cepstrum parameter is not suitable for tone recognition.Secondly,five machine learning models which are widely used in classification problems are introduced,and they are used as tone classifiers to carry out pre-comparison experiments on seven typical feature parameter sets.Then,a Chinese tone recognition algorithm based on feature fusion and random forest is proposed.In the algorithm,three fusion methods are used to optimize the feature parameters for seven typical feature parameter sets and then construct decision trees and form random forest on the three optimized fusion parameter sets.Next,carry out the model parameter optimization experiment,performance index comparison experiment and small sample training set comparison experiment in multi-person Mandarin Chinese single syllable sample data set and the experimental results are compared with other four classifiers.The experimental results show that:?The five tone classifiers can achieve the best tone recognition effect by using different fusion parameter sets respectively,indicating that when using different classifiers for tone recognition,it is necessary to specifically determine the feature parameters according to the characteristics of the model.?In the experimental results,the optimal tone recognition accuracy of the three fusion parameter sets are all above 97.50%,indicating that the three fusion parameter sets can distinguish the four tones well,and the two feature optimization methods used in fusion parameter set S2 and S3 also select the feature parameters with high distinguishability for tone recognition.?The Chinese tone recognition algorithm based on feature fusion and random forest has good tone recognition performance.The tone recognition accuracy,model generalization index(AUROC)and unbalanced data classification index(AUPRC)of random forest in the three fusion parameter sets are always maintained at more than 97.50%.Among them,the three performance indexes of random forest using the full feature fusion parameter set S1 are all higher than 98.32%.In the small sample training set,the tone recognition accuracy of random forest using three fusion parameter sets are also maintained above 93.57%.These indicate that random forest is a good tone classifier with high recognition accuracy,strong generalization ability and good stability,which has an alternative reference value for similar classification problems.