An algebraic gain estimation method to improve the performance of HMM-based speech enhancement systems

An extension to conventional Hidden Markov Model (HMM)-based speech enhancement method is developed. An algebraic method is proposed to estimate gain of speech and noise in order to improve the quality of the estimated speech. Different pronunciations and intonations may affect speech gain. Besides, gain of noise may vary remarkably from one environment to the other one. This may lead in a mismatch between energy contour of trained models and energy contour of noisy speech signal. In this work, speech gain and noise gain are estimated based on an algebraic method simultaneously in order to match gain of noisy speech and noisy model. To carry out this procedure an extension of least square method which is called non-negative least square method has been applied. Performance of the proposed enhancement method is evaluated using SNR and PESQ. Experimental results confirm advantages of this method in presence of non-stationary noise especially in lower SNR levels.