BOOSTING LOCALIZED BINARY FEATURES FOR SPEECH RECOGNITION

In a recent work, the framework of Boosted Binary Features (BBF) was proposed for ASR. In this framework, a small set of localized binary-valued features are selected using the Dis-crete Adaboost algorithm. These features are then integrated into a standard HMM-based system using either single layer perceptrons (SLP) or multilayer perceptrons (MLP). The fea-tures were found to perform significantly better (when cou-pled with SLP) and equally well (when coupled with MLP) compared to MFCC features on the TIMIT phoneme recogni-tion task. The current work presents an overview of the idea and extends it in two directions: 1) fusion of BBF with MFCC and an analysis of their complementarity, 2) scalability of the proposed features from phoneme recognition to the continu-ous speech recognition task and reusability on unseen data. Index Terms — Boosting, localized features, spectro-temporal features, speech recognition, feature fusion