Time-Domain Analysis/Synthesis of The Aperiodic Component of Speech Signals

This paper introduces a new analysis/synthesis algorithm for representing the aperiodic component of the excitation source in speech signals. This component is decomposed as a sum of random Formant Wave Forms (FWF). The time of arrivals of the FWF define the virtual excitation source. The signal is decomposed in subbands, and, according to the random modulation theory, each passband signal is represented as an envelope modulating an oscillating term. For each signal in a band, the formant filter parameters and the excitation source are estimated in the time domain. DCR tests show that this new representation scheme gives a very good fusion of the aperiodic component with the quasi-periodic component of speech. The method proposed provides new relevant parameters for manipulating the voice quality features that are linked to noise.