Some experiments on the use of one-channel noise reduction techniques with the Italian SpeechDat Car database

The use of noise reduction techniques for hands-free speech recognition in a car environment is investigated. A set of experiments was carried out using different speech enhancement algorithms based on noise estimation. In particular, linear spectral subtraction and MMSE estimators are considered with various parameter settings. Experiments were conducted on connected and isolated digits, extracted from the Italian version of the SpeechDat Car database. Recognition rates do not agree with acoustically perceived quality of noise reduction. As a result, the best performance is obtained by spectral subtraction with a suitable choice of the oversubtraction factor and a quantile noise estimator. It provides more than 30% relative performance improvement, from 94.4% of the baseline to 96.2% digit recognition accuracy.