Discriminative Spoken Term Detection with Limited Data

We study spoken term detection—the task of determining whether and where a given word or phrase appears in a given segment of speech—in the setting of limited training data. This setting is becoming increasingly important as interest grows in porting spoken term detection to multiple low resource languages and acoustic environments. We propose a discriminative algorithm that aims at maximizing the area under the receiver operating characteristic curve, often used to evaluate the performance of spoken term detection systems. We implement the approach using a set of feature functions based on multilayer perceptron classifiers of phones and articulatory features, and experiment on data drawn from the Switchboard database of conversational telephone speech. Our approach outperforms a baseline HMM-based system by a large margin across a number of training set sizes.