Using a Uniform-Weight Grammar to Model Disfluencies in Stuttered Read Speech: A Pilot Study

Stuttering is a speech disorder characterized by certain types of speech disfluencies, such as sound repetitions, which are frequent enough to be disruptive. Speech therapists frequently use manual counts of these speech disfluencies to diagnose whether a child stutters and to track improvement through a treatment program. However, these counts are subjective, inconsistent, and prone to error. We propose the use of speech recognition technology to automate these counts, thus providing an objective and consistent measurement. Since many of the disfluencies in stuttered speech obey certain regularities, we built several grammar-based language models that capture these regularities to detect disfluencies. These grammars have uniform transition weights. We tested the grammars on a short sample of stuttered speech and analyzed their performance based on word error rate and how well they detect and classify disfluencies. Results indicate that these grammars detect repetition disfluencies poorly, particularly phoneme repetitions. We are currently building a probabilistic language model and expect that the more accurate transition probabilities will lead to a decrease in false positives. We also expect that further improvement will be achieved by modeling the acoustic properties of stuttered speech.