Publication

Abstract : The nonlinear interaction of the subsystems of speech production system brings in dynamical transitions from unvoiced to voiced speech and vice versa during continuous speech production. The characterization or detection of the dynamic variation underlying the production of voiced and unvoiced speech is a fundamental step to many speech technology applications. In the present work, we study the complex behavior in the dynamics of voiced and unvoiced speech production using the framework of complex networks for the first time. The time-series corresponding to speech utterance is converted into complex network or graph using recurrence network approach to study the dynamical transitions. We find that the resultant network topology resembles the structure of the attractor representing the dynamics of the speech production system. We demonstrate the presence of either scale-free or random nature of the fluctuations from the recurrence network corresponding to unvoiced speech. Further, we show the absence of scale-free nature in the recurrence network corresponding to voiced speech. The transitions from unvoiced to voiced speech are then captured as variations of the network measures such as average clustering coefficient and characteristic path length. The performance comparison results show that the complex network measures estimated from recurrence networks provide comparable accuracy in detecting voiced/unvoiced speech with that of the state-of-the-art methods.