Machine Learning for Ambient Intelligence: Boosting in Automatic Speech Recognition

Abstract

An important aspect of Ambient Intelligence is a convenient user interface, supporting several user-friendly input modalities. Speech is one of the most natural modalities for man-machine interaction. Numerous applications in the context of Ambient Intelligence — whether referring to a single input modality or combining different ones — involve some pattern classification task. Experience shows that for building successful and reliable real life applications, advanced classification algorithms are needed providing maximal accuracy for the underlying task. In this chapter, we investigate whether a generic machine learning technique, the boosting algorithm, can successfully be applied to increase the accuracy in a ‘large-scale’ classification problem, namely large vocabulary automatic speech recognition. Specifically, we outline an approach to implement the AdaBoost.M2 algorithm for training of acoustic models in a state-of-the-art automatic speech recognizer. Detailed evaluations in a large vocabulary name recognition task show that this ‘utterance approach’ improves the best test error rates obtained with standard training paradigms. In particular, we obtain additive performance gains when combining boosting with discriminative training, one of the most powerful training algorithms in speech recognition. Our findings motivate further applications of boosting in other classification tasks relevant for Ambient Intelligence.