SNR classification based on amplitude modulation spectrogram via deep belief networks

Rebuilding noisy speech with partial spectral components which have higher SNR values brings about result of speech quality enhancement, as well as speech intelligibility improvement. Motivated by the idea of hand written digits classification with deep belief network (DBN) done by Hinton et al [8], a novel methodology is proposed here to implement SNR classification based on DBN. In our study, the whole range of SNR values is divided into subintervals, and each subinterval is considered as a class. SNR values which correspond to spectral components are represented as feature vector and are ascribed to one of the SNR subintervals. The result is helpful to tasks such as selecting noisy spectral components since each subinterval indicates a certain range of SNR values. The feature vector used in this study is an extended version of the amplitude modulation spectrogram (AMS) which was proposed by Kim et al in [7]. Experiments with IEEE corpus shows promising accuracy of classification, and reminds of the fields to be further studied.