Approach for Spectral Analysis in Detection of Selected Pronunciation Pathologies

Abstract

A framework for semi-automated detection of selected types of sigmatism is presented in this paper. A database of speech recordings was collected containing sibilant /s/ surrounded by vowels in different articulation phases. Recordings of three pronunciation modes were included into the database: normal, simulated lateral sigmatism, and simulated interdental sigmatism. The data was collected under the supervision of a speech therapy expert, who also provided labelling and annotation of each database entry. Twenty eight features of four types were extracted from each time frame within the sibilant: the mel-frequency cepstral coefficients, filter bank energies, spectral brightness, and zero-crossing rate. A feature aggregation procedure weighing the time frame location influence was proposed to describe each phoneme by a single feature vector. At the three-class classification stage, two tools were employed and compared: the random forest and support vector machine. The latter provides more accurate and repeatable classification results in each articulation phase with a median sensitivity, specificity, and accuracy exceeding 0.71, 0.85, and 0.80, respectively. The results also show that the assessment is generally more efficient when the phoneme is located at the beginning or ending of the word than when in the middle position.