A case-based approach using phonological knowledge for identifying error patterns in children's speech

Case-Based Reasoning (CBR) covers a range of different methods for organizing, retrieving, and indexing knowledge from previous cases. Thus, this methodology has been successfully applied in medical domain, due to its human and intelligent properties to diagnose the case of a patient. In the speech therapy domain, an early identification of speech sound disorders allows the diagnosis and treatment of various pathologies and may aid clinical decision-making. However, there are few proposals that use knowledge modeling for supporting speech therapists. Moreover, there is no indicative in related literature of CBR being used for detecting the phonological processes (PPs) that may occur in pronunciations. So, in this paper, we present a case-based approach that uses machine learning for predicting PPs, aiming to provide clinical support in the identification of error patterns in children's speech. The method was evaluated through a speech corpus containing near one hundred thousand audio files, collected from pronunciation assessments performed by speech-language pathologists with more than 1,000 children. Using our knowledge base along with incremental learning, we obtained an accuracy of over 93% for predicting the PPs, showing the efficiency of our method for clinical decision support.