Search

Researchers have been using signal processing based methods to assess speech from Parkinson's disease (PD) patients and identify the contrasting features in comparison to speech from healthy controls (HC). The methodologies follow conventional approach of segmenting speech over a fixed window (≈25ms to 30ms) followed by feature extraction and classification. The proposed methodology uses MFCCs extracted from pitch synchronous and fixed window (25ms) based speech segments for classification using fine Gaussian support vector machines (SVM). Three word utterances with three different vowel sounds are used for this analysis. Clustering experiments are aimed at identifying two clusters and class labels (PD/HC) are assigned based on number of participants from the respective class in the cluster. The features are divided into 9 groups based on the vowel content to evaluate the effect of different vowel sounds. Principal component analysis (PCA) is used for dimensionality reduction along with a 10-fold cross-validation. From the results, we observed that pitch synchronous segmentation yields better classification performance compared to fixed window based segmentation. The results of this analysis support our hypothesis that pitch synchronous segmentation is better suited for PD classification using connected speech.Clinical Relevance- The automatic speech analysis framework used in this analysis establishes the greater efficiency of pitch synchronous segmentation over the traditional methods.

Human speech production is a complex task that demands synchronized cognitive and muscular functioning. Assessment of a Parkinson's disease (PD) patient's speech using computational methods is a growing field of research. Existing methodologies aim at extraction and usage of features from speech to capture perturbations due to PD. In this paper, we propose a novel methodology for feature extraction and analysis. Features are extracted from each pitch cycle of the speech and variances of the features are used for analysis making this a pitch synchronous methodology. Dimensionality problem is addressed by feature selection, which is followed by an unsupervised k-means clustering to perform classification. A dataset containing 40 participants, 22 (7 female and 15 male) PD and 18 (12 female and 6 male) healthy controls (HC) is used for evaluation. The promising results yielded from this study provides support for our hypothesis that pitch synchronous speech analysis can be useful in PD analysis.