Two-channel noise reduction and post-processing for speech enhancement

Abstract

This thesis is focused on speech enhancement techniques based on short-time spectral amplitude (STSA) estimation. Some common STSA estimation methods and the-order minimum mean-square error (MMSE) estimation technique incorporating auditory masking properties (-masking in short) are reviewed. A word pair recognition accuracy test is done to assess the intelligibility of the-masking enhanced speech signals. Though the-masking technique outperforms most of the existing STSA estimation speech enhancement methods, some slight tonal distortion is audible in the processed speech signals when the background noise level is very high, because some speech spectral components are over-attenuated while some are not due to the enhancement algorithm. Two post-processing techniques are proposed to improve the quality of the single-channel-masking enhanced speech signals. One technique involves non-linear high-frequency regeneration, which uses the lower-band spectral information to re-synthesize the upper-band spectral structure. The other technique involves resynthesis of the weak spectral components using the autocorrelations of the strong spectral components. With the increasing use of mobile communication, a two-channel speech enhancement method for communication in a car environment is also studied. To achieve a better performance, the single-channel-masking method is incorporated within the two-channel enhancement system. The resulting output speech signals have low background noise and the distortion to the speech components is also very low, thus achieving an overall very satisfactory speech enhancement performance.

Novelty detection in jet engine vibration spectra

“…In [7] , the authors observed similar behaviour for background noise in acoustic signals, where narrow frequency bands used to transmit speech were seen to contain gamma-distributed background noise. In [7] , the authors observed similar behaviour for background noise in acoustic signals, where narrow frequency bands used to transmit speech were seen to contain gamma-distributed background noise.…”