In-vehicle acoustic chamber ARMA modeling and classification

Spoken dialogue interfaced systems are going to be used in vehicles to get information like navigation, near by restaurants, places of interest, etc. For these systems to perform up to the satisfaction of a user in a noisy environment like in vehicle, mainly, the accuracy of speech recognition engines should not degrade. That is, they should be able to adapt to changing environments. The changing environment can be characterized and identified by modeling and classification. This paper is focused on modeling and classifying in-vehicle acoustic chamber under different realistic operating conditions. For modeling, the auto regressive moving average (ARMA) approach is used. For classification, the support vector machine (SVM) technique is used. Real data that was collected in six different vehicles is used for both modeling and classification. The frequency response of ARMA models indicate that they depend more on the operating conditions such as window open, turn signal on, etc. than on the type of vehicle. The average operating conditions classification accuracy of 83 % is obtained after the SVM classifier was trained and tested using some of the cepstral features.