Robust Speech Activity Detection in Interactive Smart-Room Environments

In perceptive interface technologies used in smart-room environments, the determination of speech activity is one of key objectives. Due to the presence of environmental noises and reverberation, a robust Speech Activity Detection (SAD) system is required. In a previous work, a SAD system, which used Linear Discriminant Analysis-extracted features and a Decision Tree classifier, was successfully contrasted with other previously reported techniques in a set of room environment tests with the SPEECON database. In this work, the same SAD system has been tested in even more realistic conditions involving meetings, and it has been modified to significantly improve its performance. Actually, we have trained the SAD system with a subset of SPEECON data, and without any further tuning we have used it to carry out tests with the meeting databases from the NIST RT05 evaluation. In order to improve the SAD performance, we consider two additional features which are measures of energy dynamics at low and high frequencies, respectively. Besides that, two alternative classifiers have been tested, which are based on Support Vector Machines and Gaussian Mixture Models, respectively. With the latter classifier, and using both the LDA features and the low-frequency energy dynamics feature, a large improvement in speech detection performance has been observed, e.g. the NIST error rate was reduced from 20.69% to 8.47% for the RT05 evaluation data. In addition, we report the results obtained with a slightly modified version of the SAD system in the NIST RT06 evaluation.