Real-time deep learning to improve speech intelligibility in noise

点击翻译按钮获取中文摘要

Project Summary/Abstract One in eight Americans has hearing loss, and this constitutes a major health and economic burden (Blackwell et al., 2014). The primary complaint of hearing-impaired (HI) listeners is difficulty understanding speech when background noise is present (see Dillon, 2012). While hearing aids (HAs) have improved in recent years, they still provide little benefit in noisy environments. For decades, a means of improving the ability to understand speech in background noise appeared unattainable, despite substantial amounts of research by both universities and HA companies. This changed when deep learning provided the first demonstration of a single-microphone algorithm that improves intelligibly in noise for HI listeners (Healy et al., 2013, 2014, 2015). Although this algorithm provides massive intelligibility improvements (even allowing listeners to improve intelligibility from floor to ceiling levels), it is currently not implemented to operate in real time and is therefore not suitable for implementation into HAs and cochlear implants (CIs). What is needed, therefore, is a highly effective noise-reduction algorithm that is capable of operating in real time. This project aims to address this critical need. The long-term goal of the currently proposed project is to alleviate HI listeners’ predominant hearing handicap, which is difficulty understanding speech in background noise. The first aim introduces a new algorithm, based on a novel foundational scheme, that is designed to provide substantial benefit for any HI listener in real time. This algorithm will be well suited for implementation into HAs, CIs, and other face-to-face communication applications. The effectiveness of this new algorithm will be quantified using both HI and normal-hearing (NH) listeners. The second aim expands upon this new algorithm by modifying it to accept a small amount of future time-frame information, which could improve its noise-reduction performance but will introduce a brief processing delay. The rationale is that different devices have different allowable latencies. Face-to-face communication devices (HAs, CIs, etc.) have strict low-latency requirements, but other important communication systems (e.g., telephones) have different requirements. It is possible that the addition of future time-frame information within these requirements (up to 150 ms) will result in even better speech intelligibility. But the magnitude of any potential benefit is unknown. This critical information will be established currently. Using both HI and NH listeners, we will measure intelligibility for noisy sentences that have been processed using various amounts of future time information. This comprehensive fellowship training plan will provide individualized, mentored research training from world-class faculty in a highly supportive and productive environment. The proposed work will endow the applicant with the skills needed to transition to the next stage of his research career, transform our treatment of hearing loss, and substantially impact quality of life for millions of Americans.