This paper focusses on the problem of different acoustic channels in real world speech recognition scenarios. We describe an approach which enables a recognition system to compensate the influence of transmission, microphon or recording hardware. The compensation works with single-channel recordings and is based on an online estimation of the logarithmic longterm spectrum of speech. This estimated spectrum is substracted from the short-term absolute spectra. In experiments with a speaker-independent continuous speech recognizer our method essentially improves recognition performance. Using test utterances with transmission channels, which do not occur in the training material, the word error rate is reduced from 40 % to 29 %. In contrast to common compensation methods, our approach does not degrade performance, when test and training data are taken from the same channel.
Bibliographic reference. Wittmann, Matthias / Schmidbauer, Otto / Aktas, Abdulmesih (1993): "Online channel compensation for robust speech recognition", In EUROSPEECH'93, 1251-1254.