Vocal Tract Area Estimation Based On Attenuated Weighted Linear Prediction And Pathological Voice Classification

Voice is the production of a source(vocal fold)and a filter(vocal tract),and therefore,voice quality should depend on the both.As reported,vocal cords diseases are main cause of voice problems and supraglottic tract has a clear effect on the voice pathology.Existing methods that estimate the vocal tract area from vocal tract filters using speech signals suffer from inadequate elimination of the glottal wave,and the influence of non-ideal vocal tract boundary conditions.To minimize these effects on the vocal tract area estimation,we present a method that estimates the vocal tract area using closed-phase attenuated weighted linear prediction.Traditional phonetic features can achieve good results on the recognition of normal voice and pathological voice.However,there is less research on the subdivision between different kinds of pathological voice.To quantify the irregularity of the vocal tract area,we calculated the six kinds of excellent vocal tract area features over the closed glottal phase.Specific studies are shown as follows:(1)We propose a method of estimating the vocal tract area using attenuation weighted linear prediction over the entire time domain(VTA-WLP)which will not reduce the amount of analysis data,the influence of the glottis opening phase is weakened,while the role of the closed glottis phase is emphasized.We set the weight of the data points over closed glottal phase to be 1 and the data points over glottal open phase to be a positive minimum value,then we solve prediction coefficients with the smallest weighted mean square error.Finally we calculate the vocal tract area function iteratively according to the equivalent of vocal tract filter model and acoustic tube model.We verify the effectiveness of the algorithm.The vocal tract area obtained by proposed method(VTA-WLP)is more accurate than half-peak-value closed glottal phase method(VTA-HPV)comparing with MRI data with MSE_HPV=0.1542,MSE_WLP=0.0341.(2)Extracting the vocal tract parameters reflecting all kinds of abnormal voice from the vocal tract area function obtained by VTA-WLP is also one of the important contents in this study.This study is the first kind of quantifying the vocal tract irregularity to subdivide vocal cord nodules and vocal cord polyps and the best segmentation result can reach 97%,achieving an increasement of 8% compared with the feature fusion algorithm using the same database.In our proposed system,we got 99.21% accuracy with only 70 features in classifying normal voice and pathological voice,and 99.21% is one of the highest accuracies obtained so far in the literature.It provides new ideas and methods for the estimation of the vocal tract area and the identification of pathological voice.