Evaluation of Microphone Array Front-Ends for ASR - an Extension of the AURORA Framework
Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC 2004)
Abstract
The paper is focused on evaluating recognizers for automotive applications using microphone arrays. We propose to extend the framework developed within the AURORA project. Based on measurements of the impulse responses within a given car between the position of the speaker and the microphone array and based on multi-channel in-car noises recorded at the microphone array output we propose to convert the close-talk speech recordings of the AURORA databases to multi-channel databases. The resulting data can be used to evaluate algorithms of microphone array front-ends. To compare the performance of these algorithms we start with the AURORA Advanced Front-End (AFE) as a baseline for mono-channel processing and demonstrate the evaluation with basic multi-channel algorithms. Our actual goal is to propose a certain evaluation procedure and to encourage others to support it.