Extracting dynamic features using the stochastic matching pursuit algorithm for speech event detection

We extend the stochastic matching pursuit algorithm (Wang et al., 1997) to the time frequency domain to extract dynamic features of speech. Speech signals, modeled by random processes, are decomposed into a linear combination of a subset of time frequency kernels, where the random vector composed of the decomposition coefficients can be proven to capture the statistical properties of the signal. Acoustic models based on these decomposition coefficients are trained and used to locate speech events that bear distinctive signatures on the time frequency plane. A recognition paradigm based on such a speech event detection mechanism is shown to achieve comparable recognition accuracy as those of the conventional approaches based on hidden Markov models