Analysis and Interpretation of Glide Characteristics in Pursuit of an Algorithm for Recognition

In the system being developed for accessing lexical items from the speech signal, an initial step is to identify landmarks which indicate the presence of vowel, glide, and consonant segments. This thesis addresses the problem of identifying the glide landmarks. The acoustic properties measured to locate glide landmarks are based on a model of glide production. This model states that a tight constriction is formed within the vocal tract, and this narrowing causes the energy in the output speech signal to be attenuated [Bickley and Stevens, 1986). This tight constriction causes the RMS amplitude to decrease, in addition to Fl values being very low and F2 values being either extremely high (for /j/) or extremely low (for /w I). The acoustic properties measured include RMS amplitude, formant frequencies, and the transitions of these properties into and out of the glide. Previous work has located abrupt consonantal landmarks [Liu, 1995]. Further, one can expect that a reasonably error-free method can be developed for locating vowels within the speech stream. Given this information, the algorithm devised in this thesis determines whether or not a glide occurs between a consonant landmark and the following vowel by determining if the acoustic properties listed above fall into a certain range. In particular, one obtains means and covariances of glides and non-glides for the given acoustic properties from a training set, and performs hypothesis testing to determine if a given measurement in the test set is a glide or not. The thesis discusses the algorithm and the results, along with the effects of context and variability in speech and how it affects the recognition process. The overall recognition results were 88.0% for glide detection, and 90.6% for non-glide detection. Thesis Supervisor: Kenneth Noble Stevens Title: LeBel Professor of Electrical Engineering