Automatic classification of speech dysfluencies in continuous speech based on similarity measures and morphological image processing tools

Abstract Speech-language pathologists, traditionally, count the number of speech dysfluencies to measure the rate of stuttering severity. Subjective stuttering assessment is time consuming and highly dependent on clinician's experiences. The present study proposes an objective evaluation of speech dysfluencies (sounds prolongation, syllables\words\phrases repetition) in continuous speech signals. The proposed method is based on finding similarity in successive frames of speech features for sounds prolongation detection and in close segments of speech for repetition detection. Speech signals are initially parameterized to MFCC, PLP or filter bank energy feature sets. Then, similarity matrix is calculated based on similarities of all pairs of frames using cross-correlation or Euclidean criterion. Similarity matrix is considered as an image and highly similar components are extracted using proper threshold. By employing morphological image processing tools, irrelevant parts of similarity matrix are removed and dysfluent parts are detected. The effects of different feature sets and similarity measures on classification results were examined. The promising classification accuracy of 99.84%, 98.07% and 99.87% were achieved for detection of prolongation, syllable/word repetition and phrase repetition, respectively.