Unsupervised spoken keyword spotting and learning of acoustically meaningful units

Abstract

The problem of keyword spotting in audio data has been explored for many years. Typically researchers use supervised methods to train statistical models to detect keyword instances. However, such supervised methods require large quantities of annotated data that is unlikely to be available for the majority of languages in the world. This thesis addresses this lack-of-annotation problem and presents two completely unsupervised spoken keyword spotting systems that do not require any transcribed data. In the first system, a Gaussian Mixture Model is trained to label speech frames with a Gaussian posteriorgram, without any transcription information. Given several spoken samples of a keyword, a segmental dynamic time warping is used to compare the Gaussian posteriorgrams between keyword samples and test utterances. The keyword detection result is then obtained by ranking the distortion scores of all the test utterances. In the second system, to avoid the need for spoken samples, a Joint-Multigram model is used to build a mapping from the keyword text samples to the Gaussian component indices. A keyword instance in the test data can be detected by calculating the similarity score of the Gaussian component index sequences between keyword samples and test utterances. The proposed two systems are evaluated on the TIMIT and MIT Lecture corpus. The result demonstrates the viability and effectiveness of the two systems. Furthermore, encouraged by the success of using unsupervised methods to perform keyword spotting, we present some preliminary investigation on the unsupervised detection of acoustically meaningful units in speech.