Be able to train the model on the Shanon Pearce database
The suggestion is to train the model on Shanon Pearce's database. This database is interesting because it provides Binaural Room Impulses Responses, i.e. head related impulses responses measured in non-anechoic rooms, which affects the signal reaching the ear canals. The BRIRs are interesting in the way that they are different, and may be complementary, to HRIRs measured in anechoic rooms. Also, it is easy to augment the data by applying a linear windowing to the BRIRs, in order to change the slope of the reverberation of any room.
The dataset includes a set of BRIRs that were derived from a range of publicly available BRIR datasets. The BRIRs were measured using head and torso simulators (HATS) in a variety of reverberant rooms, each containing unique acoustical properties. For each room, a set of BRIRs are provided for a range of source directions around the head on the horizontal plane. The BRIRs have been equalised to remove undesired spectral colouration and to make the BRIRs compatible with diffuse-field equalised headphones. The BRIRs are provided as 2 channel WAV files with a sampling rate of 44100Hz.
some actions/tasks that might be taken
-
the dataset includes BRIRs that were derived from a range of publicly available BRIR datasets. It should be discussed with Shanon Pearce's from which database each BRIRs, which head and torso simulators are used, etc. -
the sampling directions is not uniformized: some rooms may include BRIRs for some azimuths, that may not be available to other rooms. Also, some rooms may be artificially generated from other rooms. It would be interesting to create a script that automatically generate/update the available directions for each room, or an API which returns, for example, the subset of rooms for which a list of directions is available.