Algorithms for dysfluency detection in symbolic sequences using suffix arrays

Abstract

Dysfluencies are common in spontaneous speech, but these types of events are laborious to recognize by methods used in speech recognition technologies. Speech recognition systems work well with fluent speech, but their accuracy is degraded by dysfluent events. If dysfluent events can be detected from description of their representative features before speech recognition task, statistical models could be augmented with dysfluency detector module. This work introduces our algorithm developed to extract novelty features of complex dysfluencies and derived functions for detecting pure dysfluent events. It uses statistical apparatus to analyze proposed features of complex dysfluencies in spectral domain and in symbolic sequences. With the help of Support vector machines, it performs objective assessment of MFCC features, MFCC based derived features and symbolic sequence based derived features of complex dysfluencies, where our symbolic sequence based approach increased recognition accuracy from 50.2 to 97.6 % compared to MFCC. © 2013 Springer-Verlag.