Year: 2026
Type: article
Abstract: Speech disorders can affect communication and may interfere with learning and everyday social interaction. Automatic analysis of disordered speech may provide additional information for identifying acoustic changes that are difficult to examine consistently through perceptual assessment alone. In this study, a pretrained Wav2Vec2 model was combined with an autoencoder to examine differences between control and dysarthric speech. The autoencoder was trained on speech representations obtained from the control group and reached a reconstruction error (MSE) of 0.0008. The anomaly threshold, calculated from the 99th percentile of reconstruction errors in the control samples, was 0.002143. Analysis of dysarthric speech showed differences in reconstruction error between control and affected speech segments, although the highest observed anomaly value (0.001831) remained below the selected threshold. Acoustic analysis of these segments also showed changes in formant frequencies. F1 increased by approximately 35–40%, while F2 decreased by about 20–30% compared with the reference speech samples. The recordings were analyzed in 0.5-second windows, allowing changes to be examined within individual parts of an utterance rather than at the recording level alone. These results show that Wav2Vec2 representations combined with reconstruction-based analysis can capture differences between control and dysarthric speech. Further experiments with more speakers and different speech disorders are required to determine whether the observed differences are sufficiently consistent for automatic screening and speech-monitoring applications.
Source: Вестник КазАТК
Institution: Ahmet Yesevi University
Language: English
Topic: Voice and Speech Disorders
Subfield: Speech and Hearing
Field: Health Professions
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W7225081763