Enhancement algorithm for low signal to noise ratio speech

This paper proposes a new speech enhancement algorithm to improve the quality of speech in low SNR environment. The algorithm estimates the noise spectrum in the form of prior SNR according to its Gaussian statistical model firstly. Then it estimates the prior SNR of each frequency bin utilizing the intra-frame as well as inter-frame smoothness to trace the quick change of the prior SNR. It also introduces a simple method to compute the speech presence probability to each frequency bin and obtain the new speech enhancement algorithm. Objective measurement combined with informal listening test shows that the new algorithm is more effective in eliminating musical noise and improving SNR obviously without significantly impairing the intelligibility of speech.