Bayesian changepoint detection for the automatic assessment of fluency and articulatory disorders

The accurate changepoint detection of different signal segments is a frequent challenge in a wide range of applications. With regard to speech utterances, the changepoints are related to significant spectral changes, mostly represented by the borders between two phonemes. The main aim of this study is to design a novel Bayesian autoregressive changepoint detector (BACD) and test its feasibility in the evaluation of fluency and articulatory disorders. The originality of the proposed method consists in its normalizing of a posteriori probability using Bayesian evidence and designing a recursive algorithm for reliable practice. For further evaluation of the BACD, we used data from (a) 118 people with various severity of stuttering to assess the extent of speech disfluency using a short reading passage, and (b) 24 patients with early Parkinson's disease and 22 healthy speakers for evaluation of articulation accuracy using fast syllable repetition. Subsequently, we designed two measures for each type of disorder. While speech disfluency has been related to greater distances between spectral changes, inaccurate dysarthric articulation has instead been associated with lower spectral changes. These findings have been confirmed by statistically significant differences, which were achieved in separating several degrees of disfluency and distinguishing healthy from parkinsonian speakers. In addition, a significant correlation was found between the automatic assessment of speech fluency and the judgment of human experts. In conclusion, the method proposed provides a cost-effective, easily applicable and freely available evaluation of speech disorders, as well as other areas requiring reliable techniques for changepoint detection. In a more modest scope, BACD may be used in diagnosis of disease severity, monitoring treatment, and support for therapist evaluation.