A multi-band nonlinear oscillator model for speech

Nonlinear self-oscillating systems can model speech without an external excitation that drives a conventional filter model. However, they often do not give due consideration to perceptually important but weak signal components such as the higher formants of voiced speech. To overcome this problem, we propose two frequency-domain oscillator models: a bank of sub-band oscillators with individual oscillator states and a multi-band oscillator with a single joint state vector. Their state-transition map is approximated with compactly parameterized multivariate adaptive regression splines (MARS) and the systems are successfully tested in short-term prediction and synthesis experiments with sustained vowels.