Hindi and Punjabi Continuous Speech Recognition Using CNSVM

Abstract

CNNs are playing a vital role in the field of automatic speech recognition. Most CNNs employ a softmax activation layer to minimize cross-entropy loss. This layer generates the posterior probability in object classification tasks. SVMs are also offering promising results in the field of ASR. In this article, two different approaches: CNNs and SVMs, are combined together to propose a new hybrid architecture. This model replaces the softmax layer, i.e. the last layer of CNN by SVMs to effectively deal with high dimensional features. This model should be interpreted as a special form of structured SVM and named the Convolutional Neural SVM. (CNSVM). CNSVMs incorporate the characteristics of both models which CNNs learn features from the speech signal and SVMs classify these features into corresponding text. The parameters of CNNs and SVMs are trained jointly using a sequence level max-margin and sMBR criterion. The performance achieved by CNSVM on Hindi and Punjabi speech corpus for word error rate is 13.43% and 15.86%, respectively, which is a significant improvement on CNNs.