Skip to main content

Rules

Page Updated: November 22, 2023

Added clarification regarding the use of data augmentation and data supplementation to increase the training data.

1. Teams

  • Teams must pre-register and nominate a contact person.
  • Teams can be from one or more institution.

2. Transparency

  • Anonymous entries are allowed.
  • Teams must provide a technical document of up to 2 pages describing the system/model and any external data and pre-existing tools, software and models used.
  • We will publish all technical documents (anonymous or otherwise).
  • Teams are encouraged to make their code open source.

3. What information can I use?

  • Official data: refers to the training dataset generated by using the provided scenes.tran.json file.
  • Data augmentation: refers to the use of techniques like, such as randomizing the stems, flipping right and left channels, applying SpecAugmentation, pitch shifting, etc.
  • Supplememented data: refers to the use of additional music samples. This can be by generating extra training data using MUSDB18; or the use of MedleyDB, BACH10, or FMA.

3.1. Training and development

  • There is no limit on the amount of training data that can be generated using our tools and the provided training data sets.
  • You may not use other datasets.
  • You can only use pretrained models that have been developed with public databases such as the training split of MUSDB18-HQ. You must not start with models pretrained on private datasets.
  • All the audio or metadata can be used during training and development.
  • You must not use the evaluation data set for training or tuning the system.
  • Teams that decide to use data augmentation and/or supplemented data must:
    • Using data augmentation: submit 2 systems, one without the augmentations and one with the augmentations.
    • Using supplemented data: submit 2 systems, one without the supplemented data and one with the supplemented data.
    • Using both augmentation and supplemented data: submit four systems, one only with the official data, one with the supplemented data, one with the data augmentation and one with both supplemented and data augmentation.
  • Systems using the data augmentation and/or the data supplementation will be scored and ranked in the challenge.
  • Systems trained using any other source for data supplementation not explicitly mentioned here will not enter the ranking.
Official dataOfficial data + supplementation
Official dataAll teams must submitOptional submission
Official data + augmentationOptional submissionOptional submission

3.2. Evaluation

The only data that can be used during evaluation are:

  • The audiograms giving the listener characterisation for personalisation.
  • The target gains.
  • The stereo music input signals to the hearing aid.

4. Computational restrictions

  • Systems must either be:
    • causal and low latency to allow them to work with live music, or
    • non-causal, for use with recorded music.
  • The latency restrictions for causal entries are that the output from the hearing aid at time t must not use any information from input samples more than 5 ms into the future i.e., no information from input samples >t+5 ms. See this blog post from our sister Clarity project for more.
  • There is no limit on computational cost, but entrants must report model size.
  • Teams must start with the baseline, with the blocks that can be changed labelled Enhancement
  • While HAAQI is being used for evaluation, other metrics and approaches can be used by the teams during training.

5. Submitting multiple entries

This will be allowed if different approaches are used.

6. Evaluation of systems

  • Music: we will use the MUSDB18-HQ's evaluation set which is made up of 50 songs. We will ask teams to submit defined 10 second segments from the remixed stereo for each track.
  • Gains: We will provide metadata giving the target mix for the VDBO tracks for the evaluation.
  • Listener audiograms: we will use 50 real measured audiograms that we have been collected.

Entries will be ranked according to average HAAQI score across all signals in the evaluation dataset. We will also report whether the systems are causal or non-causal in the rank order table and model size.

7. Intellectual property

The following terms apply to participation in this machine learning challenge (“Challenge”). Entrants may create original solutions, prototypes, datasets, scripts, or other content, materials, discoveries or inventions. Entrants retain ownership of all intellectual and industrial property rights (including moral rights) in and to these.

The "submission" constitutes the audio files submitted to the challenge and the accompanying technical report.

The Challenge is organised by the Challenge Organiser.

As a condition of submission, Entrant grants the Challenge Organiser, its subsidiaries, agents and partner companies, a perpetual, irrevocable, worldwide, royalty-free, and non-exclusive licence to use, reproduce, adapt, modify, publish, distribute, publicly perform, create a derivative work from, and publicly display the Submission.

Entrants provide Submissions on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE.