|
Abstract:
|
The variability inherent in speech waveforms gives rise to powerful temporal and spectral dynamics that evolve across multiple scales. As a result, optimal analysis/synthesis schemes may differ as a function of phoneme, prosody, and acoustic environment. However, most current schemes (and therefore algorithms that rely on them) operate at a fixed resolution chosen a priori. In this talk we describe methods for adaptive short-time Fourier analysis that proceed according to measures of time-frequency spread and the statistical notion of local stationarity. Subsequent synthesis is made possible through perfect-reconstruction filter bank methods and a simple modification of overlap-add procedures.
|
