Contextual keyword spotting in lecture video with deep convolutional neural network

We present a Keyword Spotting (KWS) system for lecture video that uses Deep Convolutional Neural Network (CNN) architecture. The CNN architecture enables us to use lecture video speech data that was taken under a diverse condition as a training dataset without any specialized feature. We also employ the language model and word stemming on the resulting transcript to correct misspelling and error automatically. Our model yield an average accuracy of 69.01% which is 34.53% improvement from the baseline model. We also show that an augmentation of the training dataset improves the model robustness against the testing dataset.