Aurora-5

Besides the presence of background noise and unknown frequency characteristics, the reverberation inside a room can cause an additional distortion in case of a hands-fee speech input in rooms. To cover all distortion effects as they might occur in real applications of speech recognition systems, the data base Aurora-5 has been set up in 2006. Aurora-5 is quite similar to Aurora-2. A 8 kHz version of the TIDigits is taken as basis. Special focus is put on two application scenarios. The first one is the speech input in a car where three different situations are considered. These are the usage of a close-talking microphone, the recording with a hands-free microphone and the usage of a hands-free microphone in combination with a further speech transmission over a GSM cellular network to a remote recognition system. The second scenario is the hands-free speech input in different noisy room environments, e.g. to control electronic devices like a telephone or audio/video equipment.

All noisy data of Aurora-5 have been artificially created with a tool for simulating the hands-free speech input in noisy rooms and the transmission over a cellular telephone network. A Web interface exists to experience this simulation tool . In comparison to Aurora-2 where only a subset of all available test data has been taken, Aurora-5 uses all available test data for each test set. Furthermore the variance of the distortion effects inside each test set is higher due to e.g. randomly selecting a noise segment from a whole set of noise recordings at the specific noise situation.
A set of HTK based recognition experiments is distributed with the noisy data.
A report is available containing more details about the set-up and the achieved recognition results.