Skip to main content

Listener metadata

We have provided metadata characterising the hearing abilities of the listeners, so the audio signals you generate can be individualised. The same types of data are available for training, validation and evaluation. More on hearing loss and how it is measured.

  • Training For training, we will use the same 83 audiograms from the Clarity project.

  • Validation We constructed a new set of listeners to separate the training from the validation listeners. The validation audiograms are a filtered, pseudo-random, selection from the von Gablenz and Holube (2019) dataset. Additionally, we include a "normal hearing" audiogram (L5999). Results using this audiogram can be used to evaluate the performance of systems when no amplification prescription is applied.

  • Evaluation We use a set of 52 listeners audiograms collected for the listener panel in the first Cadenza challenge. They all have two hearing aids but the hearing loss may be asymmetrical. The average hearing loss was between 25 and about 60 dB in the better ear.

Data file formats and naming conventions​

Audiogram data is stored in a JSON file per dataset with the following format.

{
    "L0001": {
        "name"               : "L0001",
        "audiogram_cfs"      : [250, 500, 1000, 2000, 3000, 4000, 6000, 8000],
        "audiogram_levels_l" : [10, 10, 20, 30, 40, 55, 55, 60],
        "audiogram_levels_r" : [ ... ],
    },
    ...
}