@inproceedings{01HNZG7EVFP92CKWDADS65RN51,
  abstract     = {{Computational auditory models have been used for decades
to develop audio signal processing algorithms in hearing
aids. Here, using a biophysically inspired auditory model in
a differentiable convolutional-neural-network (CNN)
description (CoNNear), we trained end-to-end machinelearning- (ML) based audio signal-processing algorithms
that maximally restored auditory-nerve (AN) responses
affected by cochlear synaptopathy. To this end, we used
backpropagation to develop several ML-based algorithms
that match the simulated response of the corresponding
hearing-impaired model back to the normal-hearing
response, each time using the same CNN encoder-decoder
architecture but different loss functions to achieve different
compensation of the AN responses. Evaluation of the
hearing-aid models was performed by processing sentences
of the Flemish matrix test and comparing model outcomes
with the unprocessed sentences. The magnitude spectra of all
processed sentences showed differences between the HA
models in amplification of low- and high-frequency speech
content, whereas the high-frequency processing often
introduced audible tonal distortions. Our processing showed
different enhancement of the AN population responses at the
speech onsets, vowels and consonants. We will objectively
assess the effect of the most optimal compensation
algorithms on sound quality and speech intelligibility in
future clinical experiments.}},
  author       = {{Wouters, Marjoleen and Drakopoulos, Fotios and Verhulst, Sarah}},
  booktitle    = {{Forum Acusticum 2023 : 10th Convention of the European Acoustics Association, Proceedings}},
  isbn         = {{9788888942674}},
  issn         = {{2221-3767}},
  language     = {{eng}},
  location     = {{Turin, Italy}},
  pages        = {{5}},
  title        = {{Machine-learning-based audio algorithms for hearing loss compensation}},
  url          = {{https://www.fa2023.org/}},
  year         = {{2023}},
}