Speech Signal Processing Based on Sinusoidal Models



Digital Communications This PhD project has focused on sinusoidal signal models for speech signal processing. Traditionally, the basic sinusoidal model aims at representing a time-domain speech segment as a sum of sinusoidal functions with constant amplitude and constant frequency. However, for some speech segments this constant parameter assumption is far from valid. A significant part of the PhD work dealt with generalized sinusoidal models, where the constant amplitude assumption was relaxed. In particular, models with exponential amplitude variations proved to be superior to the basic sinusoidal model, especially in transitional speech segments [J. Jensen, 2000 (a)], [J. Jensen, S. H. Jensen, E. Hansen, 2000 (b)]. It is believed that generalized sinusoidal models are particularly useful in speech transformation applications such as text-to-speech synthesis, since these models provide a "non-stationary" description of the speech production process and achieve modelled speech of transparent quality. During a research stay with Associate Professor John Hansen, Center for Spoken Language Understanding, University of Colorado at Boulder, USA, sinusoidal models were studied in a speech enhancement context. A sinusoidal model based algorithm was developed for enhancement of speech degraded by additive broadband noise. Experiments with speech signals degraded by additive white noise showed considerable improvements compared to standard enhancement algorithms. [J. Jensen, J. H. L. Hansen, 2000 (c, d)]. (Jesper Jensen, Søren Holdt Jensen)
Effektiv start/slut dato31/12/200331/12/2003