Artikel
Towards a deep learning based sound coding strategy for cochlear implants
Suche in Medline nach
Autoren
Veröffentlicht: | 12. September 2022 |
---|
Gliederung
Text
Cochlear implant (CI) users struggle to understand speech in noisy conditions. In this work, we present novel source separation architectures to improve speech in noise for CI users. These architectures range from font-end deep neural network algorithms to a novel end-to-end speech coding and denoising sound coding strategy that estimates the electrodograms from the raw audio captured by the microphone. The end-to-end speech coding strategy is a deep neural network that completely substitutes the CI sound coding strategy. We refer to this novel artificial intelligence based sound coding strategy as deepCIS when it substitutes the continuous interleaved sampling (CIS) coding strategy or deepACE when is substitutes the advanced combination encoder (ACE) coding strategy. Similar names such as deepHiRes, deepF120, deepFSP or deepCrystalis can be used for other well-known coding strategies.
The novel end-to-end deepACE approach was compared to a classic Wiener filter and the state of the art TasNet deep network architecture as front-end to the ACE. We will refer to this condition as TasNet+ACE. The different approaches were compared to assess its potential benefits in the context of CI electric hearing based on objective instrumental measures and experiments in CI users.
5 CI users participated in the experiment. Results from objective instrumental measures confirm the results observed in behavioral experiments in CI users. The behavioural experiment consisted of the HSM sentence test in quiet and with ICRA7 noise. Speech understanding results in quiet with ACE, deepACE and TasNet+ACE were equivalent. In noise, however, deepACE and TasNet+ACE significantly improved the performance of ACE. The performance with deepACE or with TasNet+Ace was not significantly different in noise, however deepACE reduces the latency of the sound coding strategy.
This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) Project ID 446611346.