Dual-channel speech enhancement with auditory spectrum based constraints

A frequency-domain, constrained iterative Wiener filtering algorithm is proposed for dual-channel speech enhancement. The iterative algorithm is shown to follow the expectation-maximization (EM) maximum likelihood estimation (MLE) technique. Constraints are applied over time and iteration on the MLE-cepstral parameters which parametrize an auditory spectrum. The constraints are adapted to changing speech characteristics over time with the aid of a speech boundary detector. The dual-channel algorithm is implemented using experimentally determined correlation functions between the two channels. Results from speech degraded by additive white Gaussian noise and actual aircraft cockpit noise are presented. Objective quality measures show improvement when compared to unconstrained dual-channel Wiener filter over a range of SNR and cross-talk levels. Time waveforms and frame-to-frame distance measures show good improvement, especially in unvoiced and transitional regions of speech. Distance measures over individual phonemes from a subset of the NIST/TIMIT database degraded with additive white Gaussian noise (AWGN) show consistent and superior quality improvement