Question
Why do some hearing aids with Deep Neural Networks have dual chips and others don’t?
Answer
First, it is important to clarify that Deep Neural Networks (DNN) can be used to solve many different problems. A DNN is trained to accurately learn complex patterns and representations from data. DNNs can be trained to perform tasks like interpretation of images and language recognition among many other applications. In hearing aids, ReSound first used a DNN to automatically recognize ears fit with hearing aids to assist users in taking a selfie of the fit of their hearing aids. This feature, Check My Fit, is housed in the Smart 3D app on the user’s smartphone and uses the processing capability of the smartphone to work.
Other ways that DNNs are currently being used relative to hearing aids include environmental recognition and classification, control of advanced features, and single channel noise reduction. Not every hearing aid that markets using DNNs is necessarily using them in the same way, but in these applications the DNN is implemented inside the hearing aids themselves.
Advancement in DNNs has happened much faster than hearing aid chip development. This means that no hearing aid manufacturer has an existing chip that is designed to accommodate the necessary architecture for DNNs or to manage the associated workloads. In particular, the workload for DNN-based noise reduction - also called “de-noising” - is problematic with existing chip technology. Therefore, we currently see two approaches in the industry:
- Combine a dedicated chip called a “DNN accelerator” with the existing sound processing chip;
- Repurpose a core in an existing chip as a DNN accelerator.
When repurposing a core in an existing chip, there is a limit to how big the DNN can be. This also limits what you can do with it. Consequently, de-noising is either severely limited or not possible, with the main benefits related to environmental classification and automatic control of advanced features.
In contrast, a dedicated DNN chip allows manufacturers to bring advanced DNN de-noising to the market now. The required architecture and computational power is handled by this DNN accelerator. Intelligent Focus with Intelligent Noise Tracker introduced in the ReSound Vivia is an example of this approach. Over the next several years, it is likely that hearing aid manufacturers will design and introduce new chips that fulfill the requirements for running DNN technology in addition to the conventional sound processing and other functionality.
