Hearing loss affects millions of people in the world, and has a deep impact on quality of life. Consequences of hearing loss can include difficulty of communication, social isolation, and even an increased risk of developing dementia. While hearing loss treatment is not restorative, hearing aids aim to compensate for the hearing loss and restore audibility of sound using a combination linear and nonlinear amplification methods.
One of the fundamental features of modern hearing aids is Wide Dynamic Range Compression (WDRC). Compression seeks to compensate for hearing loss by reducing the dynamic range of an audio signal and mapping the signal onto the remaining dynamic range of a hearing impaired individual, which is generally narrower than that of a normal hearing individual. Compression has been shown to improve loudness comfort, audibility, and intelligibility of sound for hearing impaired individuals, however, there remains significant user dissatisfaction with compressive hearing aids. This dissertation describes an end-to-end WDRC system which offers solutions to many ongoing hearing aid research challenges, and consists of the following subsystems: frequency decomposition, magnitude estimation, automatic gain control, feedback management, and user interface. Our system is open source and is part of a greater initiative to democratize hearing aid research.
Frequency decomposition is the first stage of compression. Since hearing, whether normal or impaired, is inherently frequency dependent, different frequency components must be processed differently, necessitating frequency channelization. We designed a multirate eleven band half-octave filter bank for audiometric signal decomposition. Our filter bank surpasses other non-proprietary hearing aids, offering more channels with narrower bandwidth at lower computational cost. Our novel multirate processing structure dramatically reduces the complexity of the system by processing each channel at a sampling rate proportional to its frequency. Our multirate channelizer also enables all subsequent hearing aid processing to be performed at multiple sampling rates, reducing power consumption of the entire processing chain.
The following stage of WDRC is magnitude estimation and automatic gain control. The conventional approach uses a peak detector to track a smoothed version of the signal envelope and apply gain. However, a peak detector produces an inaccurate estimate of the signal envelope with ripple. Moreover, the effects of attack and release times in hearing aids are not well explored in academic literature. We derived a closed-form solution relating attack and release time to other hearing aid parameters. Based on this derivation, we designed a highly accurate envelope detector based on the Hilbert Transform, followed by a precise automatic-gain-control algorithm yielding highly accurate dynamic performance.
Central to the automatic-gain-control algorithm is the input-output transfer function, which maps input signal magnitude to output magnitude. The conventional WDRC input-output transfer function is a piece-wise segmented function, defined by four or more parameters. Conventional WDRC also overamplifies low-level signals, which reduces output signal-to-noise ratio and causes the hearing aid to sound noisy. We designed a novel curvilinear transfer function for WDRC which simplifies the parameterization of compression, reduces distortion caused by hard knee points, and improves signal-to-noise ratio by reducing amplification of low-level noisy.
Lastly, a major difficulty in hearing aid research is acoustic feedback management. Adaptive filters are highly effective at cancelling feedback, however, there is a delay in their response to a change of acoustic conditions, such as a hand touching the ear, and adaptive filters may cancel useful components of the input signal as well as feedback. We designed a feedback cancellation algorithm which uses variable frequency shifting to successfully eliminate feedback in realistic hearing aid test conditions.
