| Abstract |
: |
Although many signal-typing studies have been published, they are primarily based on manual inspection and experts’ judgments of voice samples’ acoustic content. Software may be required to automatically and objectively classify pathological voices into the four signal types and to facilitate experts’ opinion formation by providing specific signal type determination criteria. This paper suggests the coefficient of normalized skewness variation (CSV), coefficient of normalized kurtosis variation (CKV), and bicoherence value (BV) based on the linear predictive coding (LPC) residual to categorize voice signals. Its objective is to improve the performances of acoustic parameters such as jitter, shimmer, and the signal-to-noise ratio (SNR) in signal type classification. In this study, the classification and regression tree (CART) was used to estimate the performances of the acoustic, CSV, CKV, and BV parameters by using the LPC residual. In the investigation of acoustic parameters such as jitter, shimmer, and the SNR, the optimal tree generated by jitter alone yielded an average accuracy of 78.6%. When the acoustic, CSV, CKV, and BV parameters together were used to generate the decision tree, the average accuracy was 82.1%. In this case, the optimal tree formed by jitter and the BV effectively discriminated between the signal types. To perform accurate acoustic pathological voice analysis, signal type quantification is of great interest. Automatic pathological voice classification can be an important objective tool as the signal type can be numerically measured. Future investigations will incorporate multiple pathological data in classification methods to improve their performance and implement more reliable detectors. |
