This thesis studies transform-domain model-based algorithms to improve the feedback and noise control performance of hearing aids processing wideband speech in non-stationary environments. Subband adaptive filter structures are investigated for continuous and non-continuous adaptation feedback compensation, and the particle filter framework is used to develop state-space model-based speech enhancement algorithms.
Subband feedback compensation is shown to offer several advantages. The flexibility offered by the frequency division allows subband systems to offer faster and more stable convergence for wideband signals, and better tracking of changing acoustic feedback paths. Robustness is also improved, as divergence caused by path changes or input signal correlation is confined to individual frequency bands.
A Rao-Blackwellized particle filter (RBPF) algorithm is proposed for enhancement of speech discrete cosine transform (DCT) coefficients, and is evaluated in comparison to the standard fullband RBPF. The DCT subband decomposition is shown to enable improved modeling of wideband speech signals, especially in spectral troughs, thereby decreasing intra-speech noise in both best-case and real-world conditions.
A novel particle filter algorithm is also proposed for short-time spectral amplitude (STSA) speech enhancement. A dynamic model of spectral amplitude evolution is used to allow for speech signal correlation in the frequency domain. Two variants of the basic algorithm are presented: the first incorporates phase information to improve the spectral amplitude estimates; the second uses interacting multiple models to account for speech presence uncertainty. The monaural algorithm is extended to the binaural case with a filter based speech and noise parameter estimator. The estimator exploits knowledge of the diffuse noise field coherence to separate the clean speech and noise power spectra without external noise or voice-activity estimation.
In both feedback and noise control, the transform domain allows for processing strategies that take into account the diverse frequency-dependent characteristics of the wideband signals. Incorporating models of speech, noise and the acoustic environment allows the addition of time-domain constraints and a-priori knowledge to address signal non-stationarity. The developed algorithms are evaluated using real speech and noise signals recorded with a commercial hearing aid.