Research On Pathological Voice Repair Algorithms

Speech is the most direct way for humans to transmit information.In recent years,organic or neurological lesions of the vocal organs make the incidence of voice diseases higher and higher.The ambiguity of voice will affect people’s communication efficiency.Therefore,using non-invasive signal processing to repair pathological voice can improve the quality of damaged voice and provide better services for human society.In addition,pathological voice repair has wide application prospects in the fields of speech recognition,machine navigation,speech enhancement,speech communication system,military criminal investigation and covert surveillance.Aiming at improving the intelligibility of pathological voice,a complete system of pathological voice repair is designed and implemented in this paper,which has completed the following tasks:(1)Analyze the principle of the vocal system to provide theoretical basis for the study of pathological voice pathogenesis.Establish the mathematical model of the whole vocal system and introduce the excitation model,vocal tract model and radiation model in detail.In view of polyp voice signals with high incidence,the acoustic characteristics are analyzed from time domain and frequency domain.In addition,the acoustic features are classified and summarized from both short-term and long-term aspects in order to analyze the pathological voice signals more efficiently.Study the algorithm to improve the recognition rate of pathological voice.Aiming at the low recognition rate of pathological voice by traditional acoustic features,a pathological voice recognition algorithm based on E-BLSP feature is proposed.E-BLSP and other traditional features(LPCC,MFCC)are input into SVM and DNN networks to study the recognition performance of single feature and feature combination under two classifiers.Accuracy,ROC and other indicators are used to objectively evaluate the effectiveness of the proposed algorithm.(2)Study the excitation parameters reflecting the prosody characteristics in the excitation model.Aiming at the invalidation of the pitch frequency extraction of pathological voice by traditional algorithms,an algorithm based on wavelet transform and HHT transform is proposed to accurately extract pitch information of pathological voice.Study the relationship between vocal tract model parameters,vowel type and voice timbre.Aiming at the deviation and instability of the formants of pathological voice,the pathological voice is iteratively repaired based on LSP features by referring to the characteristics of normal voice features to reconstruct the vocal tract characteristics of pathological voice.Study speech synthesis algorithms based on different synthesis rules.Based on the repaired pitch frequency parameters and vocal tract parameters,a linear predictive parameter synthesis method is selected in this paper.Study the indicators to evaluate the speech quality.The performance of the pathological voice repair system is evaluated from three aspects: time domain,frequency domain and auditory domain.The experimental results show that the pathological voice repair system achieves satisfactory results in terms of speech intelligibility.