Aurora-3

Aurora-3 consists of noisy speech data as recorded as part of the SpeechDatCar project inside cars. These data have been used for the evaluations inside the Aurora group to reflect a small vocabulary recognition of real recordings in a noisy environment. 5 subsets from 5 different languages were available at the time of the evaluation process. Different sites volunteered to define the experiments and provide a set of scripts for performing the recognition. As with Aurora-2 HTK has been taken as basis for the recognition experiments. All 5 data sets including the scripts and configuration files are available from ELRA/ELDA. A report is available for each individual experiment.