Deep neural network quantizers outperforming continuous speech recognition systems

Abstract

In Automatic Speech Recognition (ASR), the acoustic model (AM) is modeled by a Deep Neural Network (DNN). The DNN learns a posterior probability in a supervised fashion utilizing input features and ground-truth labels. Current approaches combine a DNN with a Hidden Markov Model (HMM) in a hybrid approach, which achieved good results in the last years. Similar approaches using a discrete version, hence a Discrete Hidden Markov Model (DHMM), have been disregarded in recent past. Our approach revisits the idea of a discrete system, more precisely the so-called Deep Neural Network Quantizer (DNNQ), demonstrating how a DNNQ is created and trained. We introduce a novel approach to train a DNNQ in a supervised fashion with an arbitrary output layer size even though suitable target values are not available. The proposed method provides a mapping function exploiting fixed ground-truth labels. Consequently, we are able to apply a frame-based cross entropy (CE) training. Our experiments demonstrate that the DNNQ reduces the Word Error Rate (WER) by 17.6 % on monophones and by 2.2 % on triphones, respectively, compared to a continuous HMM-Gaussian Mixture Model (GMM) system.