Study On Speech Intelligibility Enhancement In Low Signal-to-Noise Ratio Environment

Speech enhancement technologies improve the quality or intelligibility of speech via eliminating noise from speech signal contaminated by various noises. Existing speech enhancement algorithms suppress noise as much as possible to improve the speech quality, i.e., the auditory comfort. Little attention has been paid to the speech intelligibility improvement when people design speech enhancement algorithms. Improving the speech intelligibility, however, has more application value in low signal to noise (SNR) environment compared with the speech quality improvement. This thesis focuses on speech enhancement technologies aiming to improve the speech intelligibility. The main research work is as follows:First, we proposed an amplitude spectrum distortion controlling based speech enhancement algorithm. In low SNR environment, the estimation error of the prior SNR and the amplifying distortion of the gain function degrade the speech intelligibility. To reduce the negative impact on intelligibility, the estimation value of the prior SNR is amended. And the gain function value is controlled in a constraint range. In addition, in order to judge the time-frequency areas which are harmful to the speech intelligibility, both the distortion criterion based on speech amplitude and the distortion criterion based on noise amplitude are derived and used to decide whether the time-frequency areas can be constrained. Experimental results in usual noise environment of various SNR show the effectiveness of the proposed algorithm in the aspect of improving speech intelligibility.Second, we proposed a subspace speech enhancement algorithm based on joint distortion controlling. In traditional subspace based speech enhancement algorithms, only the speech distortion is considered to be minimized when deriving the speech estimator. Similar to the speech distortion, the residual noise also exists in the enhanced speech and has negative impact on speech intelligibility. With the amplitude amplification distortion constraint, two speech estimators are derived based on minimizing speech distortion and residual noise, respectively. And then we can make a weighted sum of these two estimators to get the final estimator. Experimental results show that the intelligibility of the speech enhanced by the proposed algorithm outperforms that of the unprocessed noisy speech and that of the traditional generalized subspace methods.