ESCA Tutorial and Research Workshop on
This paper focusses on the evaluation of the segmentation and labelling component of large vocabulary automatic speech recognition systems, drawing on experience from the Alvey Speech Demonstrator project. We distinguish between intrinsic evaluation, directed at judging the output of a component in its own terms, and extrinsic evaluation, directed at judging how well the output of one component will serve as input to the next. We discuss two methods for intrinsic evaluation, one based on segment overlap and one on tick-level label equality, and two methods for extrinsic evaluation, one based on a heuristic path computation and one on constrained DTW.
Bibliographic reference. Thompson, Henry S. (1989): "Evaluation of phoneme lattices: four methods compared", In SIOA-1989, Vol.2, 131-134.