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 recording a constrained space given by an HNM decoding. With these overall project goals, the primary research efforts and results over the last quarter have included: (1) an extensive literature survey of research adaptation; (2) development of a trigram word prediction tool for the use in experiments t6 estimate the entropy of conversational English; (3) further experimental exploration of dependence tree topology design and extension of the modeling framework to handle continuous observation vectors; (4) initiated work on HMM topology design; and (5) furthered efforts on establishing a baseline HTK recognition system for a task of recognizing the Marcophone natura numbers data, on which we currently achieve 76% word accuracy.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 30, 1995
- Accession Number
- ADA314528
Entities
People
- Mari Ostendorf
Organizations
- Boston University