PATHOLOGICAL VOICE DETECTION USING TURBULENT SPEECH SEGMENTS

Identification of voice pathologies using only the voice signal has a great advantage over the conventional methods, such as laryngoscopy, since they enable a non-invasive diagnosis. The first studies in this area were based on the analysis of sustained vowel sounds. More recently, there are studies that extend the analysis to continuous speech, achieving similar or better results. All these studies use of a pitch detector algorithm to select only the voiced parts of the acoustic signal. However, the existence of a pathology affecting the speaker’s vocal folds produces a more irregular vibration pattern and, consequently, a degradation of the voice quality with less voiced segments. Thus, by selecting only clear voiced segments for the classifier, useful pathological information may be disregarded. In this study we propose a new approach that enables the classification of voice pathology by also analyzing the unvoiced information of continuous speech. The signal frames are divided in turbulent/non-turbulent, instead of voice/non-voiced. The results show that useful information is indeed present in turbulent or near unvoiced segments. A comparison with systems that use the entire signal or only the non-turbulent frames shows that the unvoiced or highly turbulent speech segments contain useful pathological information.