Artificial neural network categorization of speech fluency measures across speech disorders

By A.P. Salvatore, M. Cannito, A Biswas, R. Ingham, B.K. Bender, and B Manriquez

This project investigated the patterns of speech fluency behavior found in five normal, five spasmodic dysphonic, and five stuttering speakers through the application of unsupervised, self-organizing artificial neural-network (ANN) techniques. This investigation developed an evolving ANN model to describe the relationships present in speech samples based upon a reading of the Rainbow Passage. Historically, clinical judgment of voice and speech status is based upon training provided under the tutelage of an expert, experience, and statistical data. Training acquired under the tutelage of an expert has the potential for bias. Also the use of statistical methods may not be appropriate for assessing multidimensional, dynamic and non-linear data such as speech fluency measures. The identification of common speech fluency patterns across different clinical diagnostic groups of patients may be instructive in developing efficacious clinical training materials and procedures. The current on-going investigation indicates that a trained neural network can accurately categorize speech samples collected from the above three groups of participants. This investigation provides insights into the nature of normal and pathological voice/speech disfluency, and the identification of subtypes of speakers not previously appreciated.