Hearing-aid algorithms stop prescribing gain from the audiogram alone: closed-loop training compensates hair-cell and nerve damage separately, reaching HASPIv2 0.92 in babble

The training target of the hearing-aid processor shifts from executing prescribed gain rules to making the impaired auditory model's output match the normal-hearing model's output as closely as possible, so compensation can be designed separately for outer-hair-cell gain loss and for cochlear synaptopathy, the loss of synapses between inner hair cells and auditory nerve fibers. Conventional hearing aids rely on frequency-specific gain from the audiogram, wide dynamic-range compression, and fixed attack and release times; these assumptions only partly compensate outer-hair-cell gain loss and cannot correct the complex distortions caused by cochlear and neural damage. On a single fictive mild high-frequency sloping audiogram, transfer functions measured with 5-second white noise show the outer-hair-cell algorithm giving more gain to low-level sounds than high-level sounds, giving more gain at higher than lower frequencies in line with the audiogram shape, and giving more gain at 2–3 kHz than NAL-NL2; the NAL-NL2 reference was configured in openMHA with fast compression and a 52 dB SPL wideband compression threshold.

The synaptopathy-compensating algorithm provides no gain and instead attenuates the sound by about 5–7 dB while altering the temporal structure of the speech waveform, consistent with the view that driving the remaining auditory nerve fibers with strong inputs may not help. The outer-hair-cell algorithm takes an amplification route, whereas the synaptopathy algorithm takes a temporal-waveform route, and both emerge from the same closed-loop framework. In recordings with a head-and-torso simulator surrounded by 8 loudspeakers, target speech at 70 dB SPL and babble noise at −3 dB SNR, the synaptopathy algorithm scored 0.53 on HASPIv2, close to NAL-NL2 at 0.53, while unprocessed speech scored 0.46; HASPIv2 accounts only for audiogram-related frequency selectivity and compression loss, not synaptopathy.

When outer-hair-cell and synaptopathy damage are trained together, the algorithm supplies high-frequency gain to restore audibility but less than the outer-hair-cell-only version, and its waveform falls between the two. Combined damage is not a simple sum of the two strategies; closed-loop optimization yields a compromise between restoring audibility and driving the remaining auditory nerve. Under the same babble condition, the combined algorithm scored 0.86 on HASPIv2, above NAL-NL2 at 0.53 and below the outer-hair-cell version at 0.92; at the auditory-nerve level, both the synaptopathy and combined algorithms narrowed the gap between impaired and normal models, especially at low characteristic frequencies, but could not fully restore the auditory-nerve population response.

The demonstration is limited to a single fictive mild high-frequency sloping audiogram and covers only outer-hair-cell gain loss and cochlear synaptopathy, excluding inner-hair-cell cilia damage and stria vascularis metabolic changes; evaluation is offline, and HASPIv2 is a model prediction rather than a patient measurement.

Synaptopathy compensation is trained only on an extreme profile with a single high-spontaneous-rate fiber per characteristic frequency, which the authors describe as a test of maximal compensation potential; the human phase-locking limit is assumed at 1.5–1.6 kHz and remains debated; the model uses future context, so real-time use requires balancing hardware latency against future context, and the authors note total delay generally needs to stay below 10 ms.