Simulating an acoustic recognizer

In this paper we describe a speech recognition simulator whose input is phonemic transcriptions and which produces scores that simulate the scores generated by a real acoustic recognizer. We used a phonemic distance metric to approximate acoustic distances. A dynamic programming algorithm was used to find the optimum match between phoneme strings based on a distinctive feature metric. Taking the log of phonemic scores and taking word length into account improved performance. Mapping the distribution of acoustic distances of a large set of word pairs onto the phonemic distances for the same word pairs enabled us to add a random noise factor to simulate acoustic variability. By adjusting the simulated acoustic variability, we can match the performance characteristics of different acoustic recognizers.