SBIR-STTR Award

This is a proposal for the development of a system for continuous speech recognition of moderate size vocabularies with a low perplexity grammar, which would achieve robust, high performance under severe operational conditions. The objective is to provide advanced recognition capabilities for tactical environments, such as military aircraft or battle management, in which speakers are exposed to high ambient noise, and subject to psychological and physical stress. The main efforts in the development will consist in elaborating and testing new methods of handling stress and noise conditions, integrating the improvements into the existing HMM continuous speech recognition system under development at dragon, and designing an actual prototype. Speech recognition capability in real military environments. Speech interface to command and control battle management systems. Small vocabulary voice data entry applications in noisy environments.

The aim of this work is to develope a speaker-independent small-vocabulary continuous speech recognizer using computationally efficient algorithms that is robust to a range of types of back-ground noise and to voice changes induced by stress and noise. Only single-sensor methods will be considered in this work. The performance of current Dragon technology will form an initial benchmark against which performance can be measured. Sub-word and whole-word modeling will be considered, as well as possible combinations of the two. A variety of established and novel techniques for coping with noise and voice changes will be explored. When simulation experiments have permitted a suitable set of techniques and parameter values to be determined, they will be implemented in a real-time demonstrator. Anticipated

Benefits:
The technology to be developed here should increase the feasbility of using speech recognition in a variety of militarily and commercially important environments including aricraft, factories, tanks and road vehicels. Continuous speech recognition makes for faster input; and speaker-independence allows delay-free switching between users. Much of what is learned shoudl be transferable to large-vocabulary systems.

Keywords:
Noise-Robust Continuous Speech Recognition Speaker-Independent Small Vocabulary Real-Time ...