Regularized sparse features for noisy speech enhancement using deep neural networks

A speech enhancement algorithm improves the perceptual aspects of a speech degraded by noise signals. We propose a phase-aware deep neural network (DNN) using the regularized sparse features for speech enhancement. A regularized sparse decomposition is applied to noisy speech and the obtained sparse features are combined with robust acoustic features to train DNN. Two time-frequency masks including ideal ratio mask (IRM) and ideal binary mask (IBM) are estimated. An intelligibility improvement filter is applied as post-processer to further improve the intelligibility. During waveform reconstruction, the estimated phase is used for better quality. The results show that the proposed algorithm achieves better speech intelligibility and quality. Besides, less residual noise and speech distortion is observed. By using the TIMIT and LibriSpeech databases, the proposed algorithm improved the intelligibility and quality by 14.61% and 42.11% over the noisy speech.