Phone-mediated word alignment for speech recognition evaluation

The authors present an algorithm that produces a word-based alignment of two text strings by postprocessing the output of a phone-based alignment procedure. The alignment is performed using an unconstrained endpoint optimization, and minimizes the readjustment of word boundaries as determined by the phone-level alignment. Although the word-level recognition performance summaries were not strikingly different from those of similar word-based alignment procedures, the actual word alignments from the phone-mediated system were better. The present level of performance of the system has proved to be acceptable for a variety of applications, including evaluation of speech recognition system performance