Speech enhancement using auditory-based spectral amplitude estimators

Speech enhancement improves the quality of speech by removing certain amount of noise from noisy speech while keeping the speech component undistorted. In this project, a new family of Bayesian estimators has been proposed such that its cost function includes both a power law and a weighting factor. The cost function parameters are chosen based on the characteristics of human auditory system and the estimator gain is made frequency dependent. The estimator gain decreases at high frequencies, improving the noise reduction and reducing the speech distortion. The original speech signal is subjected to three types of noises, namely, white noise, pink noise and cockpit noise. The proposed estimator is applied to the noisy speech signal to obtain the enhanced speech signal and the objective measures of speech quality such as the segmental SNR (SNRseg) and the log-likelihood-ratio (LLR) are obtained. Performance of the weighted P-SA estimator is verified and it was found that the new estimator achieves better enhancement performance than existing Bayesian estimator such as those based on the minimum mean-square error(MMSE) of the short-time spectral amplitude (STSA), the MMSE of the logarithm of the STSA(LSA) or the weighted Euclidean (WE) error.Master of Science (Signal Processing