Deep learning restores speech intelligibility in multi-talker interference for cochlear implant users

Agudemu Borjigin ^1,2,3, Kostas Kokkinakis ^4,5, Hari M Bharadwaj ^6,7,8

⁰Weldon School of Biomedical Engineering, Purdue University, West Lafayette, 47907, IN, USA. dagu@wisc.edu.

Scientific Reports +

|

June 9, 2024

View abstract on PubMed

Summary

Deep neural network (DNN) algorithms like RNN and SepFormer significantly improve speech understanding for cochlear implant (CI) users in noisy conditions. These advanced algorithms effectively reduce both stationary and non-stationary noise, outperforming traditional methods.

Area Of Science

Audiology
Artificial Intelligence
Signal Processing

Background

Cochlear implants (CIs) struggle in noisy environments due to limitations of current noise reduction algorithms.
Conventional algorithms are ineffective against non-stationary noise, such as multi-talker interference.
Deep neural networks (DNNs) show promise for speech enhancement but require testing with CI users.

Purpose Of The Study

To evaluate the effectiveness of two DNN algorithms, Recurrent Neural Network (RNN) and SepFormer, for improving speech intelligibility in cochlear implant users.
To assess the performance of these DNNs in reducing both stationary and non-stationary noise.
To compare DNN-based noise reduction with conventional methods for CI users.

Main Methods

Implemented RNN and SepFormer DNN algorithms for speech audio processing.
Trained the algorithms using a customized dataset of approximately 30 hours.
Tested the algorithms with thirteen adult cochlear implant listeners in various noise conditions.

Main Results

Both RNN and SepFormer significantly enhanced speech intelligibility for CI listeners in noise.
The DNN algorithms improved performance in stationary non-speech noise.
Substantial improvements were observed in non-stationary noise conditions, outperforming conventional strategies.

Conclusions

DNN algorithms, specifically RNN and SepFormer, offer a promising solution for enhancing speech understanding in noisy environments for cochlear implant users.
These advanced algorithms effectively address the limitations of traditional noise reduction techniques, particularly in complex, non-stationary noise.
The findings suggest a significant advancement in addressing listening challenges faced by individuals with cochlear implants.