HearAdvisor's AI model beats gold standard in hearing aid speech clarity

HearAdvisor, a platform providing audio-performance metrics for hearing aids, has released a massive dataset and a new AI model that predicts how well listeners with hearing loss understand speech through commercial devices. The dataset originates from a blind online listening test (MUSHRA-inspired) where website visitors rated recordings from 83 commercial hearing aids across 72 acoustic scenes on a five-point 'Ease of Understanding' scale. After quality screening, 104,298 ratings remained. To predict these subjective scores, the researchers fed aided audio and a clean speech reference through a frozen Whisper encoder, subtracted the internal representations, and trained a small MLP head on that difference embedding.

The resulting model significantly outperforms HASPIv2, the current standard metric for intelligibility. At the scene level, the learned metric achieved a correlation of r=0.92 with listener ratings versus HASPIv2's r=0.83. Gains were especially strong in loud environments (r=0.89 vs 0.75) and quiet ones (r=0.79 vs 0.58). In loud scenes, the model's performance matched the split-half reliability of the human raters, while in quiet it approached that ceiling. The model also correctly responded to controlled changes in gain and signal-to-noise ratio. Together, this dataset and model provide a more realistic, listener-focused way to evaluate hearing aid speech enhancement, potentially helping consumers make better-informed purchasing decisions.