Acoustic and perceptual studies of Lombard speech: application to isolated-words automatic speech recognition

The purpose of this study was (1) to determine what are the acoustic-phonetic differences between speech produced in quiet and speech produced in noise (Lombard speech) and (2) to evaluate the influence of these differences on human listeners and automatic speech recognizers. The acoustical analyses, done at the phonetic level on about 40 parameters, showed significant differences in variability for male and female speakers. In addition to replicating previous studies, the authors investigated more parameters, and examined the influence of the Lombard effect on female speakers. Perceptual experiments, run for foreign listeners, exhibited a decrease of the intelligibility for some confusable subsets of the vocabulary studied. The findings are correlated with the performance of a DTW-based recognizer, and it is observed that recognition scores cannot necessarily be improved by reducing the variability for one specific parameter. It is also found that recognition scores are not directly related to the increase of the vocal effort and cannot be predicted from speech variability