HS101 - Data-Driven Deep Neural Network Noise Reduction

Abstract:
This presentation will describe a data-driven deep neural network noise reduction and how it differs from conventional signal model-based noise reduction techniques. A study evaluating the data-driven noise reduction and its benefits to patients against a signal model-based noise reduction will be shared. Study results reveal a strong preference for the data-driven noise reduction over the signal model-based noise reduction, with participants favoring it 89% of the time when the strongest noise reduction is applied. These results highlight the potential of data-driven solutions, suggesting that data-driven noise reduction techniques may significantly benefit patients over conventional methods.

Summary:

Rationale
Artificial Intelligence (AI) is rapidly evolving our world today. This technological wave will have a profound impact on the audiology industry and for both audiologists and patients. One example is in AI based single channel noise reduction techniques. With this data-driven method, a Deep Neural Network comprised of nodes that are analogous to neural connections in the brain is trained to identify speech and noise. It is trained to recognize different speech e.g. different languages and male and female speech as well as a lot of different types of both stationary and non-stationary noises. The data-driven deep neural network noise reduction applies what it has learned to suppress noise during hearing aid use. Data-driven deep neural network noise reduction differs from signal model-based noise reduction in that it recognizes speech and noise whereas signal model-based noise reduction must make some assumptions on the speech and noise signals, e.g. the noise is assumed to be changing slower than the speech in waveform. The data-driven noise reduction is expected to provide more efficient reduction of noise as it can identify many noises instead of making assumptions.

Objective/Purpose
The objective of the study was to assess the application of a data-driven deep neural network noise reduction and investigate patient preferences for the technology. The hypothesis was that there would be a patient preference for the data-driven deep neural network-based noise reduction over a signal model-based noise reduction.

Methods
Twenty adult experienced hearing aid users participated in the study. Recordings were made through a pair of test hearing aids fit to a head and torso simulator (HATS) with occluding domes for two different types of single channel noise reductions at varying strength levels: the data-driven noise reduction and the signal model-based noise reduction. The noise reduction was combined with 4-microphone binaural beamforming in each case. The HATS was placed within a quadrophonic speaker array where realistic sound scenes were played. The loudspeaker configuration for the sound scenes was supplemented with a talker originating from a fifth loudspeaker in front the HATS. The signal-to-noise-ratio of the talker relative to the background noise was between +3 to +5dB. The recordings were presented to the participants following the Senselab protocol from FORCE Technology. The participants performed a forced choice comparison task where they listened to six different sound scenes with the types and levels of noise reduction as conditions and indicated which they preferred in each comparison. The results were analyzed using a binomial statistical test.

Results
Significant preferences for the data-driven noise reduction compared to the signal model-based noise reduction were demonstrated in all conditions. When the strongest levels of noise reduction were applied, the test participants preferred the data-driven reduction 89% of the time (p < 0.001).

Conclusion
Data-driven noise reduction can provide a preferred listening experience for patients compared to conventional signal model-based noise reduction techniques.

Brief Summary of Clinical Takeaways:
The clinical takeaway for this session is that data-driven noise reduction in hearing aids offers patient benefits, but also trade-offs. It can carry extra costs and compromises, as additional processing capabilities than are typical for hearing aids are required. This presentation will help audiologists to understand the benefits offered by current technology. With this knowledge, they will be better equipped to make appropriate recommendations and guide patient expectations to today’s noise reduction technology.