Round 1: 2021
A critical problem for a hearing aid wearer is to pick out speech when there is background noise. In the first Enhancement Challenge, we invite you to find new ways to process the sound within a hearing aid, so that it’s easier for a hearing-impaired listener to hear and understand the speech produced by their conversational partner (the “target talker”).
Round one is about the intelligibility of speech in a home environment with a single target talker and a single noise source. The scenario is as follows:
- A home environment: a simulated cuboid room with low reverberation.
- One static target talker producing speech drawn from a new corpus of 10,000 sentences produced by 40 British English speakers.
- Background noise from a single static source: either competing speech or a typical domestic sound such as the whirr of a washing machine.
- The processing is to be personalised for specified hearing-impaired listeners.
All approaches are welcome. We are providing a suite of open-source tools within a baseline system for you to improve on. These include
- A tool for generating realistic training materials.
- A baseline model of a hearing aid for you to better!
- A model of hearing loss.
- A model that estimates binaural speech intelligibility.
There are cash prizes for the systems that top our evaluation. See the rules.
1st February 2021
- Release of documentation and description of baseline system.
- Example dataset published.
15th February 2021
- Full dataset
- Scene generation code
- Baseline system.
1st June 2021
Evaluation data released.
15th June 2021
More details of the challenge
Please post questions on the Google Group discussions board.