Query-By-Example Spoken Term Detection Using Generative Adversarial Network

Several Neural Network (NN)-based representation techniques have already been proposed for Query-by-Example Spoken Term Detection (QbE-STD) task. The recent advancement in Generative Adversarial Network (GAN) for several speech technology applications, motivated us to explore the GAN in QbE-STD. In this work, we propose to exploit GAN with the regularized cross-entropy loss, and develop a framework featuring GAN, trained using Gaussian Mixture Model (GMM)-based posterior labels. The proposed GAN maps the speech-specific features to the unsupervised posterior labels. This mapping represents the speech through an unsupervised GAN posteriorgram (uGAN-PG), for matching the query (keyword) with the utterances in the document. The QbE-STD, using the proposed posteriorgram is performed on the TIMIT database. We compare the performance of the proposed uGAN-PG with the unsupervised Deep Neural Network (DNN) posteriorgram (uDNN-PG) and obtained the relative performance improvement of 10.32 % Mean Average Precision and 5.6 % Precision by considering top N queries (p@N) over uDNN-PG.