An objective parallel evaluator of segmentation/Classification performance for multiple systems

A major difficulty in the development of methodologies for segmentation and classification in automatic recognition of continuous speech is the determination of objective, reliable performance statistics. Compounding this difficulty is the large amount of data necessary to make reasonably accurate performance estimates. The system to be described provides for concurrent objective evaluation of up to five independent segmentation/classification methods against a single, carefully transcribed referent. A basic assumption of the evaluator is that the systems to be compared, as well as the referent, can each use the same digital data as input. Violation of this assumption would lead to time-shift errors, and objective comparison among systems would be exceedingly difficult. For segmentation, the evaluator provides first-order statistics, at the phonetic, class and summary levels, in the form of highly concise tables for the following four types of errors: 1) Missed events; 2) Adventitious events; 3) Misplaced events; and 4) Adventitious and misplaced events. For classification, first-order statistics are derived in the form of confusion matrices at the phonetic, class and summary levels. While the system is still in the developmental process, it is operational and currently used. Examples of output will be presented.