Aurora 2 based on TIdigits downsampled to 8kHz. There are three different test sets of noises:

1. Set A: Subway, Babble, Car, Exhibition Hall
2. Set b: Restaurant, Street, Airport, Train Station
3. Set c: Subway, Street (with different frequency characteristics)

The training conditions on Aurora 2 database are of two kinds:

1. Training on clean data
2. Multi-condition Training on noisy data search).

The number of training sentences are 8440 and there are 1001 test sentences per test set.

We used three Front-End configurations for our baseline experinments (click on each one to see the baseline results):

• HTK
• WI007
• WI008

Aurora 3 Database

Five European languages constitute the Aurora 3 database:

• Finnish
• Spanish
• German
• Danish
• Italian

All the data have been recorded in three noisy conditions: quiet, low noisy (low), high noisy (high).

There are three experimental setups:

• Well-Matched (WM)
  • 70% of all utterances in quiet, low, and high conditions were used for training, and the remaining 30% for testing

• Medium Mismatched (MM)
  • 100% hands-free recordings from quiet and low for training and 100% hands-free recordings from high for testing

• High Mismatched (HM)
  • 70% of close-talking recordings from all noise conditions for training and 30% of hands-free recordings from low and high for testing

Click here to see the Aurora 3 baseline results.

Aurora 4 Database

Aurora 4 database is based on the WSJ0 collection, contained of 7138 training data.

Six different noises have been artificially added: Car, Babble, Restaurant, Street, Airport, and Train Station.

Training data sets are of three kinds: clean, multicondition, and noisy.

There are 14 test sets of two sizes: small (166 utterances), and large (330 utterances).

You can find more information about the Aurora 4 database and the imported Lattices in here.

Click here to see the Aurora 4 baseline results.