Noise adaptive deep belief network for robust speech features extraction

Deep neural networks (DNNs) have been widely used for acoustic modeling and also feature extraction in automatic speech recognition (ASR) systems. In the acoustic modeling, due to wide speaker variations, some methods have been proposed for adapting DNN to new speakers. In this paper, we propose to use this idea for adaptation of deep belief network (DBN), as a spectral feature extractor, to unknown and new noisy conditions. To this end, we use DBN in two approaches. In the first approach, we use DBN to extract logarithm of Mel sub-band energies (LMFBs) from noisy speech spectrum as robust features. In the second approach, we use DBN to estimate clean speech spectrum from the noisy one. Then, we propose to add a linear layer to DBN (as a linear adaptation method) in order to adapt DBN to unknown and new noisy conditions. This method has been applied to both mentioned DBNs. This linear layer can be added to DBN input or its output. Experimental results on Aurora2 database for ASR show that the proposed adaptation methods increase DBN performance for extracting more robust speech features. Where we have 7.5% and 11% improvement in recognition accuracy by adapting DBNs which are used as LMFB extractor and clean speech spectrum estimator, respectively.