《SUBMITTED TO IEEE TRANS. ON AUDIO,》.pdfVIP

SUBMITTED TO IEEE TRANS. ON AUDIO, SPEECH, AND LANGUAGE PROCESSING 1 Acoustic Modeling using Deep Belief Networks Abdel-rahman Mohamed, George E. Dahl, and Geoffrey Hinton Abstract—Gaussian mixture models are currently the domi- using a feature vector that describes segments of the temporal nant technique for modeling the emission distribution of hidden evolution of critical-band spectral densities within a single Markov models for speech recognition. We show that better critical band. Sub-word posterior probabilities are estimated phone recognition on the TIMIT dataset can be achieved by replacing Gaussian mixture models by deep neural networks using feedforward neural networks for each critical band and that contain many layers of features and a very large number these probabilities are merged to produce the ?nal estimate of parameters. These networks are ?rst pre-trained as a multi- of the posterior probabilities using another feedforward neural layer generative model of a window of spectral feature vectors network. In [8], the split temporal context system is introduced without making use of any discriminative information. Once the which modi?es the TRAP system by including, in the middle generative pre-training has designed the features, we perform discriminative ?ne-tuning using backpropagation to adjust the layer of the system, splits over time as well as over frequency features slightly to make them better at predicting a probability bands before the ?nal merger neural network. distribution over the states of monophone hidden Markov models. Feedforward neural networks offer several potential advan- tages over GMMs: Index Terms—Acoustic Modeling, deep belief networks, neural ? Their estimation of the posterior probabilities of HMM networks, phone recognition states does not require detailed assumptions about