Automatic text alignment for speech system evaluation

This paper describes an algorithm which automatically aligns two text strings, maximizing the phoneme-to-phoneme similarities between the strings, The alignment problem is formulated in an unconstrained endpoint dynamic optimization framework, and operates on the phonemic transcriptions of the input text strings. A modification of the standard dynamic programming approach allows insertion and rejection errors in either string to be correctly identified. Benchmark evaluations show the performance of the algorithm approaches "the expert speech scientist." The algorithm is an ideal tool for automated evaluation of experimental data, including data collected from speech intelligibility experiments and speech recognition experiments.