Automatic Fluency Assessment by Signal-Level Measurement of Spontaneous Speech

In its narrow sense, the term fluency connotes fluidity of speech. This study is a step in the quest for objective language assessment methods one of which is rating for oral fluency in a second language. In particular, we seek to find what measures obtained from a spontaneous utterance can be used as predictors of fluency and, to assess the utility of a set of acoustic measures obtained by signal-level measurements towards predicting fluency automatically. Experiments done using an ESL data set of spontaneous speech show that articulation rate and phonation-time ratio are good predictors of fluency, in line with earlier findings. Our contribution is to use signal-level measurements as quantifiers of perceived fluency in a logistic regression framework and to show the existence of an alternate approach to ASR-based fluency assessment, which, owing to unacceptable levels of recognition accuracies, have limited use in real testing scenarios. Our results have implications in developing fluency assessment systems for language-resource scarce settings as well as for a wide variety of testing scenarios.