Research On Speaker Verification And Parkinson Disease Diagnosis Based On Voice Feature Learning

Speech feature is an important biological marker, which has been widely used in speaker recognition, disease diagnosis and other fields. Among them, the speaker verification and diagnosis of Parkinson disease are the hot and difficult points in the application of speech feature in recent years. Although much positive progress has been made in the related research, there still are some problems in feature compression,sample selection, classifier optimization etc, and these aspects are important for improving the quality of speech feature learning.Under the support of relevant funds, this paper aimed to solve the two specific application problems of the speaker verification and diagnosis of Parkinson's disease,studied the relevant procedures including speech feature compression, sample selection,classifier optimization and proposed three algorithms, thereby achieving obvious improvement. The main related works are as follows:(1) proposed a multiple-type speech feature wrapper evolutionary selection framework algorithm. At the same time, based on the framework algorithm, this paper proposed a specific implementation method-four types of speech feature wrapper genetic selection algorithm, which shows the concrete implementation example of the framework algorithm. The proposed algorithm is helpful to offering standardization of research of relevant methods.(2) proposed speaker verification algorithm based on multiple-type speech feature learning in a very low SNR environment. Firstly, extraction method of multiple-type voice features under low SNR is proposed, especially the pitch frequency extraction.Secondly, feature selection classification ensemble model based on CAGA(Chain-like Agent Genetic Algorithm) and GMM-UBM is designed, through feature selection on the extracted features four feature fusion for the speaker recognition. The proposed algorithm realized speaker verification algorithm under extreme low SNR environment and is helpful to solving the existing problem of extracted features of low quality.(3) proposed Parkinson's disease diagnosis algorithm based on speech sample multi-edit nearest neighbor algorithm and random forest achieve the optimization of speech samples through speech sample multi-edit nearest neighbor algorithm; realize ensemble learning of the samples by introducing random forest to. This algorithm realizes the optimization of speech samples, thereby improving the quality of thesamples. The algorithm provides new research ideas and methods reference for the current relevant researches.The studies provide method reference for the related researchers and new idea for improving pattern classification research, based on speech features learning and have important research value.