Article Preview
Top1. Introduction
The process of mapping spoken utterances into the corresponding text is known as speech recognition. The emergence of CNN-based acoustic models improves the performance of speech recognition task significantly (Abdel-Hamid et al., 2012, 2013; Passricha & Aggarwal, 2018; Sainath et al., 2013; Singhal et al., 2018). CNNs efficiently model the temporal correlation with adaptive content information. In most of the CNNs, the last layer of the architecture is the softmax layer that is used for classification purpose. SVMs also offer good recognition rate in speech recognition task (Shi-Xiong & Gales, 2013). The basic principle of SVMs is maximizing the margin between two classes. The classification in which whole data points are classified into two-classes is known as linear classification and such SVM as linear SVM (Vapnik, 1995). Its extended version that is used for sequence recognition is known as structured SVM. SVM has many unique features: first, it maximizes the margin between the classes and it minimizes an upper bound of generalization errors. Second, SVM guarantees to have a global optimal solution because SVM’s nature is convex for optimization problems. Third, the size of SVM can easily be determined by calculating the number of support vectors that are trained during training. However, SVMs have some shortcomings like there is a need to predefine a feature space, and it does not deal well with arbitrary temporal dependencies and variable length data sequences.
In continuous speech recognition task, some training points are more important than others. Thus, high priority should be assigned to useful training points and low priority should be given to other training points. SVMs have less effect of outliers and noise so they effectively classify the acoustic features. Inspired by the advantages of both and to overcome the limitations of SVMs, a new model named as convolutional neural support vector machine (CNSVM) is proposed by combining CNNs and SVMs. In this model, SVMs simply replace the last layer of CNNs i.e. the softmax layer. Earlier, Suykens and Vandewalle (1999) proposed MLPs and SVMs as a combined model where SVM optimizes the weights of MLP. Nagi et al. (2012) proposed a similar model convolutional neural support vector machines for multi-robot systems. Instead of training each component separately or jointly, they trained their model using a new training approach. They allowed the model for computationally light incremental learning based on small chunks of training samples. It is tested for finger-count 0 to 5 from finger-count gestures. However, CNSVMs differ from old ones because, in this model, parameters are trained jointly, not incrementally. Agarap (2017) also proposed an architecture by combining CNN and SVM for image classification. This architecture was trained using SGD training method and achieved a test accuracy of 90.72% while the CNN-softmax achieved a test accuracy of 91.86% on popular image recognition dataset MNIST.
Veselý et al. (2013) investigate the sequence discriminative training of neural networks and find that minimum Bayes risk (MBR) criterion performs better than maximum mutual information (MMI). sequence level max-margin (sMM) and state-level minimum Bayes risk (sMBR) are different algorithms that train the parameters of CNNs and SVMs jointly. In this work, the softmax layer is replaced by SVMs and its advantages are explored. The proposed model uses CNNs to learn the feature space from the speech signal. CNNs also handles the temporal dependencies. SVMs perform the classification of the acoustic frame. In CNSVMs, the scores from the SVMs replace the frame-level posterior probability. CNSVM is evaluated on Hindi and Punjabi speech datasets for word error rate (WER) and it achieves 13.43% and 15.86% WER for Hindi and Punjabi speech respectively. Experimental results show that the CNSVMs offer a relative reduction of 4.84% and 4.28% in WER when compared to CNN on Hindi and Punjabi speech datasets respectively.