What is DNN and How Does It Empower Users of ReSound Vivia Hearing Aids?

Question

Answer

A deep neural network (DNN) is a component of AI machine learning that is utilized in many ways across countless fields and settings. In audiology, DNNs are currently used to transform hearing aids by enhancing sound clarity and auditory experiences. Yet, as with all up-and-coming technological advancements, DNNs face several challenges, including issues related to its power consumption in a miniature digital electronic device and the need for optimized training that contains valid and unbiased datasets. Addressing these challenges is key in maximizing the effectiveness of DNNs and ensuring the successful implementation of deep learning in real-world scenarios.

But what exactly IS DNN, and what are some of the ways it is used?

Deep learning is a subset of machine learning that employs multilayered networks and algorithms that are not designed by human engineers; the algorithms learn using artificial neural networks inspired by human biological neuroscience.¹ These artificial neural networks are composed of multiple layers between the input and the output, known as “hidden layers.” This is where the “deep” concept is derived. DNN is characterized by having at least three hidden layers between the input and output systems, but there can be hundreds or even thousands of layers. Each layer builds upon itself to create meaningful representations and categorizations of data.² When trained with an appropriately large dataset, the computer becomes better at processing due to increased opportunities for learning.

In the field of audiology, DNNs in hearing aids are designed to improve auditory experiences for users. This includes reading soundscapes to produce clearer versions of sounds, organizing and balancing auditory inputs, isolating speech from noise, and removing background noise to deliver cleaner and clearer sound. ReSound Vivia uses DNN to classify and remove noise from the environment. ReSound DNN effectively denoises the auditory environment, which enables the user to hear speech more easily in difficult listening environments.

The use and training of a DNN to achieve the desired outcome must be judicious and balanced. Overtraining the DNN system can result in reduced generalizability to novel listening environments that are slightly different than how the system was trained. On the other hand, undertraining can lead to errors in the outcome. Another important issue is the quality of the data being fed to the DNN and how the model interprets it, which can pose a significant challenge in deploying DNNs in real-world situations.³ Additionally, the computation time required for training DNNs can be resource-intensive and time-consuming. All these factors need to be weighed to find the optimal amount of training, confidence in the integrity of the data trained, and conservation of power needed to run the operation. After all, this system must be designed to work within the confines of a hearing aid that is worn unobtrusively for the human ear.

To responsibly and accurately employ DNN to provide maximal benefit for hearing aid users, ReSound applies the most up-to-date knowledge of DNNs when implementing this technology in Vivia hearing aids. The system is powerful enough to perform 4.9 trillion operations per day, requiring the DNN architecture to be optimized for efficiency to ensure a full day of battery power for the hearing aid user. For accuracy without over- or undergeneralization of unique auditory environments, ReSound DNN was trained on 13.5 million real-life conversations in noise.  This allows for ReSound Vivia hearing aids to intelligently classify and reduce noise in the listening environment, within a microRIE model that provides all-day battery runtime.

Despite these technological advances, it is important to recognize that no hearing aid can always predict which speaker the user wants to hear in a noisy multi-talker environment. To address this concern, ReSound uses DNN in combination with beamforming directionality and has designed the technology to prioritize auditory awareness, spatial hearing, and localization benefits. The user is kept in control of who they want to hear by turning or looking at the speaker of interest, since the beamforming directional response improves the signal-to-noise ratio (SNR) for a narrow beamwidth in front of them. This is especially useful in a common real-world situation where the user is having a meal at a crowded restaurant and wants to focus on the conversation across the table rather than the distracting, sometimes louder conversations of other people seated nearby. At the same time, spatial hearing benefits are available via multiband directionality that prioritizes SNR benefit while preserving interaural listening cues that are essential for the localization of other sounds in the environment. In short, ReSound uses DNN with beamforming directionality to empower the user to be in control of their hearing.

When implementing new technological advantages, ReSound strives to explore the most advantageous ways to support the unique needs of hearing aid users. With a conscientious approach to DNN that keeps the user in the driver’s seat of what they want to hear, bundled into a discreet hearing aid that provides a full day of battery runtime, the ReSound Vivia is designed to empower the hearing aid user for optimal hearing benefits no matter where the day takes them.

References

1. LeCun, Yann & Bengio, Y. & Hinton, Geoffrey. (2015). Deep Learning. Nature. 521. 436-44. 10.1038/nature14539.
2. Kotsiopoulos, T., Sarigiannidis, P., Ioannidis, D., & Tzovaras, D. (2021). Machine learning and deep learning in smart manufacturing: The smart grid paradigm. Computer Science Review, 40, 100341. https://doi.org/10.1016/j.cosrev.2020.100341
3. Khoei, T. T., Ould Slimane, H., & Kaabouch, N. (2023). Deep learning: Systematic review, models, challenges, and research directions. Artificial Intelligence Review. Retrieved from https://link.springer.com/article/10.1007/s00521-023-08957-4