Convolutional Feature based Vision Transformer Model for Speech Command Recognition

Speech command recognition (SCR) is the process of recognising the short speech commands uttered by speakers using statistical and deep learning (DL) based models. With emerging developments in the field of DL research, deep convolutional neural network (DCNN) models have appeared as a better alternative over the conventional statistical models. However, in recent years vision transformer models (ViTs) have shown superior performance over DCNN for image classification tasks. Moreover, in ViTs the number of trainable parameters are comparatively less than DCNN models. Therefore, in this work we have proposed an improved ViT model incorporated with convolutional features for the speech command recognition task. The motivation behind the incorporation of ViT model is that it extracts the information of formant distribution from small patches of the input mel-spectrogram. Whereas, the convolutional blocks incorporated to the ViT model captures the prominent local features more effectively. The proposed convolutional feature based improved ViT model is trained and tested with the google speech commands v2 (SCV2) dataset. From the experimental findings it is observed that the proposed improved ViT model performed better than the existing ViT model in terms of accuracy for the SCR task.