A Pipeline Consisting of Pattern Recognition and Finite Automata for Recognizing VCV Productions in the Study of Vocal Hyperfunction

Relative fundamental frequency (RFF) is an acoustic measure used to quantify vocal effort in voice science. Since it seeks to capture transitions between (i.e. to/from) steady-state vowels and unvoiced consonants, any machine learning approach to recognize patterns in these transitions should require time properties capable of identifying the sequence of phonemes. At the same time, Neural Networks (NN) have become a ubiquitous solution for data-driven problems, and Recursive NNs (RNN) provide a time-series schema to address time-dependent problems. Indeed, typical Neural Network solutions require either a time-series schema like in RNN or some spectral transformation to be able to handle time-dependent data. In this study, we decided to ignore - at least momentarily - any time-series dependency of the data and employed a simple NN to classify elements of the speech. Later, a State-Machine was used to identify their sequence with the purpose of localizing the transitions between voiced and unvoiced sounds in vowel-consonant-vowel (VCV) productions. The goal of this study was to demonstrate that a pipeline consisting of time-agnostic (Neural Network) and time-dependent (State Machine) components can be used to recognize time-dependent patterns in VCV productions.