Investigation of a Novel Class of Compact Oscillator Models for Speech

The classical nonlinear oscillator (NO) model has been reported to result in more natural-sounding speech than traditional linear methods. In this paper, we investigate the application of a novel class of oscillator models called Interval Domain (ID) models to speech. Unlike the classical NO model, ID models have compact representations that can describe a wide range of linear and nonlinear signals with only a limited number of parameters. We show that low-order ID models suciently describe certain nonlinear behaviors, such as quadratic phase coupling. The relationship between model parameters and fundamental frequency is also explored. Pitch estimates can be obtained from ID models at very low signal-to-noise ratios, suggesting that this relationship is fairly robust. Finally, we demonstrate that these models are capable of preserving the fundamental frequency and waveform morphology of speech. The spectral envelope, however, is not always retained. All experiments are conducted using sentences from the TIMIT database.