Acoustic Analysis for Detection of Different Voice Pathological Conditions
Voice is a vital thing for humans. Affected or stricken voice will impact his/her mental & social existence entirely. Acoustic analysis is one of the non-invasive methods to evaluate the physical properties of speech.Our objective is to propose an automatic tool to classify different speech pathological conditions using acoustic features. VOICED database is used and it contains voice recordings of sustained vowel phonation /a/ with duration of 5 seconds. Framing of a signal into 16 segments and acoustic feature extraction foreach speech signal was performed on Time, Frequency, Cepstral and Statistical domains. Under these 4 major domains, there are 28 features extracted as acoustic parameters. To choose the best features, a feature selection method was utilized by Correlation Coefficient analysis. 10 prominent features were selected by feature selection method to improve model performance. Random forest classifier gave the model accuracy of 85% on binary class and 81% on multiclass classification of speech signals.
Keywords - Signal Processing, Speech Signal Processing, Pathology Detection, Acoustical Analysis, Feature Selection & Machine Learning.