The 3rd Clarity Prediction Challenge: A Machine Learning Challenge for Hearing aid Speech Intelligibility Prediction

This paper reports on the design and outcomes of the third Clarity Prediction Challenge (CPC3) on predicting the intelligibility of hearing-aid processed speech. CPC3 is the final round in a series of UKRI-funded challenges to advance intelligibility prediction models. Building on CPC1 (2022) and CPC2 (2023), it introduces a larger dataset of intelligibility scores from listeners who have hearing loss, additional hearing-aid processors, and more complex acoustic scenes. Unlike previous rounds, the evaluation set contained entirely unseen scenes and algorithms, providing a stronger test of generalisation. This paper describes the publicly available dataset, challenge tasks, and outcomes. CPC3 attracted 21 systems from 15 teams. Results show that reference-free systems based on pre-trained transformer models, which had shown promise in CPC2, perform strongly under the more demanding CPC3 conditions. The best system achieved a Root Mean Square Error (RMSE) of 24.98% beating the Hearing Aid Speech Perception Index (HASPI) baseline by 4.49% absolute. We also examine the complementarity among top systems; failure cases including listener outliers, and implications for future challenge designs and evaluations.