High-Order Modeling Techniques for Continuous Speech Recognition.
Abstract
This research aims to develop new and more accurate stochastic models for speaker independent continuous speech recognition by developing acoustic and language models aimed at representing high-order statistical dependencies within and across utterances, including speaker, channel, and topic characteristics. These techniques, which have high computational costs because of the large search space associated with higher order models, are made feasible through a multi-pass search strategy that involves rescoring a constrained space given by an HMM decoding. With these overall project goals, the primary research efforts and results over the last quarter have included: (1) developed much of the theory for two new models for adaptation, (2) further explored methods for robust dependence tree topology design and implemented training algorithms for hidden dependence tree models; (3) repeated sentence-level mixture language modeling experiments with new versions of NAB training set, showing improvements in both perplexity and word error rates; (4) developed software tools for using HTK in experiments on HMM topology design; and (5) furthered efforts on establishing a baseline HTK recognition system
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 31, 1995
- Accession Number
- ADA314529
Entities
People
- Mari Ostendorf
Organizations
- Boston University