Detection and Classification of Stuttering from Text

Stuttering is a language disorder that disrupts the flow of speech, resulting in considerable challenges for those affected by it. Activities such as making phone calls and using new technologies like ASR (Automatic Speech Recognition), can pose a major challenge for someone with stuttering, often making it difficult or even impossible for them to use these tools. Speech-language pathologists and speech therapists manually count the number of stuttering instances in a person’s speech to assess the severity of stutter and track its progress over time. However, this method is time-consuming, costly, and prone to errors. Therefore, the automatic evaluation of stuttered speech is crucial for automating the counting and classification of different types of stuttering and is valuable for clinical assessment. In general, this research consists of two parts: detecting stuttered words, and classifying types of stuttering in the DISCO text dataset. The proposed model, CNN-Seq-GANBERT, initially identifies stuttered words with $\mathbf{9 8. 6 \%}$ accuracy. The network’s output is subsequently processed by a CNN classifier to distinguish between different stuttering types. The classifier achieved an accuracy of $\mathbf{9 7. 9 2 \%}$ on the training data, while the best accuracy on the test data was $\mathbf{8 0 \%}$. Finally, considering that this research is the first to perform text classification on this dataset, the results obtained were compared with the implementation of another method using GRU and transformer neural networks, as well as two machine learning methods, SVM and KNN, to validate the performance of the proposed model.