Speech enhancement using a truncated and constrained minimum variance estimator in non-uniform wavelet filterbanks.

摘要

In this dissertation, the new algorithm, Truncated and Constrained Minimum Variance Estimator (TCMVE), for speech enhancement is developed with the design of critical filter banks. The design of critical filter banks using Perceptual Wavelets (PW) and UnDecimated Wavelet Packets (UDWP) are extended and generalized. Furthermore, the TCMVE is developed as a new method for noise suppression, which can be used in any subbands. The developed TCMVE provides a unified way based on the constraint to design several estimators such as Minimum Variance Estimator (MVE) and Spectral Domain Constrained Estimator (SDCE). It is shown that the MVE and SDCE become the same estimator for the particular constrained condition of TCMVE. The combination of TCMVE/MVE with critical subbands designed by PW and UDWP provides effective methods to remove non-white noise, such as pink and babble noise, as well as white noise, in terms of Automatic Speech Recognition (ASR) and Segmental Signal to Noise Ratio (SSNR).