AIR Database

Database of room impulse responses for the development and evaluation of dereverberation algorithms

The Aachen Impulse Response (AIR) database is a set of impulse responses
that were measured in a wide variety of rooms. The initial aim of the
AIR database was to allow for realistic studies of signal processing
algorithms in reverberant environments with a special focus on hearing
aids applications.

The first version was published in 2009 and offers binaural room impulse
responses (BRIR) measured with a dummy head in different locations with
different acoustical properties, such as reverberation time and room
volume. Besides the evaluation of dereverberation algorithms and perceptual
investigations of reverberant speech, this part of the database allows for
the investigation of head shadowing influence since all recordings where
made with and without the dummy head.

In a first update, the database was extended to BRIRs with various azimuth
angles between head and desired source. This further allows to investigate
(binaural) direction-of-arrival (DOA) algorithms as well as the influence
of signal processing algorithms on the binaural cues.

Since dereverberation can also be applied to telephone speech, the latest
extension includes (dual-channel) impulse responses between the artificial
mouth of a dummy head and a mock-up phone. The measurements were carried
out in compliance with the ITU standards for both the hand-held and the
hands-free position.

For the latest extension, the IND has carried out new measurements of binaural room impulse responses in the Aula Carolina Aachen. The former church with a ground area of 570m² and a high ceiling shows very strong reverberation effects.

The database will successively extended to further application scenarios.
Please send us comments and feedback if you use the AIR database.
Suggestions for further measurements are also very welcome.

All impulse responses of the AIR database are stored as double-precision
binary floating-point MAT-files which can be directly imported into
MATLAB. Additionally, a load function (load_air.m) as well as an example
script (load_air_example.m) is provided to allow for a rapid integration
into existing evaluation frameworks.