ACID/HNN: a framework for hierarchical connectionist acoustic modeling

We propose the ACID/HNN framework for context dependent large vocabulary conversational speech recognition (LVCSR) using connectionist acoustic models. Our approach advocates the principles of modularity and hierarchy for the estimation of thousands of context dependent posterior HMM state probabilities. We argue that a hierarchical organization of the acoustic model is crucial in obtaining competitive performance with connectionist estimators. We introduce ACID, an Agglomerative Clustering scheme based on information divergence and use it to induce soft decision trees for hierarchical classification. A Hierarchy of Neural Networks (HNN) is then applied to the estimation of conditional posterior probabilities. We discuss the benefits of hierarchically structured acoustic models for speaker adaptation and scoring speed-up. Finally, we present experiments on the Switchboard conversational telephone speech corpus, currently a major focus of research in the LVCSR community.