| Abstract [eng] |
In this research techniques are analyzed and new methods are proposed for non-invasive voice pathology detection. Such voice signal parameters as Mel-frequency cepstral coefficients, autocorrelation, shape of signal envelope and patient questionnaire data such as age, sex, smoking, voice quality self evaluation and other were used to improve classification accuracy of voice pathology detection. New data dependent random forest technique and association rules based classification methods were proposed for voice signal parameters and questionnaire data classification. Classification results of voice signal parameters and questionnaire data separately and together revealed, that highest classification accuracy can be received by using both types of data together. Experiments also showed that voice pathology can be successfully detected using only questionnaire data. Proposed data dependent random forest technique allowed to achieve statistically significant classification accuracy improvement. |
