DYSFLUENCY CLASSIFICATION IN STUTTERED SPEECH USING DEEP LEARNING FOR REAL-TIME APPLICATIONS

Stuttering detection and classification are important issues in speech therapy as they could help therapists track the progression of patients' dysfluencies. This is also an important tool for technology-assisted speech therapy. In this paper, we combine MFCC and phoneme probabilities to train a neural network for stuttering detection and classification of four dysfluency types. We evaluate our system on the UCLASS, FluencyBank and SEP-28K datasets and show that our system is effective and suitable for real-time applications.