On the characteristics and effects of loudness during utterance production in continuous speech recognition

We have checked out that, in speech recognition based telephone applications, the loudness with which the speech signal is produced is a source of degradation of the word accuracy if it is lower or higher than normal. For this reason, we have carried out a research work which has reached three goals: (a) get a better understanding of the speech production loudness (SPL) phenomenon, (b) find out the parameters of the speech recognizer that are the most affected by loudness variations, and (c) compute the effects of SPL and whispery speech in large vocabulary continuous speech recognition (LVCSR). In this paper we report the results of this study for three different loudnesses (low, normal and high) and whispery speech. We also report the word accuracy degradation of a continuous speech recognition system when the speech production loudness is different than normal as well as the degradation for whispery speech. The study has been done using the TRESVOL Spanish database, that was designed to study, evaluate and compensate the effect of loudness and whispery speech in LVCSR systems