Clarity: machine learning challenges for improving hearing aid processing of speech in noise

Author: Dr Simone Graetzer, Acoustics Research Group, University of Salford, UK

Background: In recent years, rapid advances in speech technology have been made possible by machine learning challenges such as CHiME and Hurricane. In the Clarity project, the machine learning approach is applied to the problem of hearing aid processing of speech-in-noise. In the first round, the scenario is a simulated cuboid-shaped living room in which there is a single static listener, target speaker and interferer. The entrants’ tasks are to improve the target speech intelligibility at the listener’s ear and to predict it. The software provided includes a generative tool and baseline hearing aid, hearing loss and speech intelligibility models. The databases provided include a large speech database. This is the first machine learning challenge to consider the problem of hearing aid speech signal processing.

Method: The target is within +/- 30 degrees azimuth of the listener, inclusive. The interferer is an omnidirectional point source (speech or domestic noise) at the same elevation. Signals are processed by the baseline hearing aid model, a configuration of the openMHA system for a simple Behind-The Ear (BTE) model. Subsequently, signals are passed through the baseline hearing loss model and the speech intelligibility model (Modified Binaural Short-Time Objective Intelligibility or MBSTOI). Challenge entrants replace the baseline hearing aid or the hearing loss and/or speech intelligibility models.

Results: Baseline system performance measured for the development dataset as MBSTOI values as a proxy for measured intelligibility scores varies according to, for example, SNR and distances between the listener and target.

Conclusion: The first round of the Clarity challenges aims to produce improved hearing aid algorithms for speech signal processing. The first enhancement challenge involves developing hearing aid models that improve on the baseline, while the first prediction challenge involves improving on the baseline speech intelligibility prediction models.