Unsupervised keyword spotting using bounded generalized Gaussian mixture model with ICA

In this paper, bounded generalized Gaussian mixture model (BGGMM) using independent component analysis (ICA) is proposed and applied to an existing unsupervised keyword spotting setting for the generation of posteriorgrams. The ICA mixture model is trained without any transcription information to generate the posteriorgrams which further labels the speech frames of the keyword example(s) and test data. For the detection of occurrence of a specific keyword in the test data, the posteriorgrams of one or more keyword examples are compared with the posteriorgrams of test utterances using the segmental dynamic time warping (DTW). A score fusion method is used to obtain the result of the keyword detection by ranking the distortion scores of all the test utterances. The TIMIT speech corpus is used for the evaluation of this unsupervised keyword spotting setting. The keyword detection results demonstrate the viability and effectiveness of the proposed algorithm in unsupervised keyword spotting framework.