A novel approach to unsupervised pattern discovery in speech using Convolutional Neural Network

Abstract In this paper, a novel approach to unsupervised pattern discovery for speech signals is proposed. Recently, we introduced an image processing method (IPM) that extracts the desired keywords present in a pair of speech utterances. This method performs well in detecting true positives but, at the same time it also detects higher number of false positives. Therefore, this paper aims to reduce the detection of false positives and improve the accuracy of the pattern discovery task. In the proposed work, we use the Convolutional Neural Network (CNN) as a binary classifier to detect the true and false keyword match candidates. A new frame histogram technique is introduced to generate sufficient training samples from IPM to train the CNN. The trained CNN model classifies the matched patterns into true and false classes and identifies the pairs of speech documents that contain the same keyword. The proposed method is evaluated on the Hindi as well as Bengali speech databases. The results are compared with state-of-the-art methods. The detected matched pairs of speech utterances are grouped into broader domain clusters using Newman’s clustering algorithm. These clusters are useful for speech retrieval tasks.