Nonlinear state space embedding features and their application to robust speech segmentation

The aim of our research is to develop features from a state space embedded signal useful for high resolution and robust real-time speech segmentation. Since the set of state-space trajectories of a system can completely describe the system, state space embedding of a signal is typically used to study qualitatively any nonlinearities of the system generating a signal. However. while it may be easy for one to observe patterns in the state space trajectories of a system, it is often difficult to quantify what is observed. Two novel features are extracted from a state space embedded signal using concepts from differential geometry. These features are computed iteratively on the one-dimensional speech signal and they completely characterize the state space trajectories formed by the signal. The results obtained show that these features are particularly useful for classifying voicing states and can detect these phoneme boundaries with a resolution of four samples.