AI in Hearing Aids

thank you for sharing thr infomation, yes different brands’ AI feature are different, the best way to find the one which suitable for the users is try them all !

Based on public info from GN and Starkey, models like ReSound Vivia and Starkey Edge AI use deep neural networks (DNNs) for speech enhancement and noise suppression. GN has mentioned that Vivia’s model has ~3.9 million parameters, and it’s reasonable to expect Edge AI to be in a similar multi-million-parameter range.

These aren’t Large Language Models (LLMs). They’re compact audio DNNs, often built with convolutional and/or recurrent layers, trained on real-world audio to separate speech from noise in real time. With quantization and careful engineering, the working memory needed for inference is on the order of a few megabytes, which today’s ultra-low-power hearing-aid chips can handle. Modern chip packaging makes it feasible to combine an ultra lower power MCU and a powerful DSP to run audio DNN, and stack a few MB of SRAM to fit the AI model, all in a very small footprint and within tight latency and battery constraints.

So yes, it’s AI, just not generative AI. LLMs typically have billions of parameters and need tens of GB of memory, which is a completely different class of problem from running an on-ear DNN for hearing enhancement.

I am not at all an expert in this topic, but am I right to conclude that AI is not at all active in HAs but is merely a tool that is being used to produce smarter sound processing algorithms in HA chips? The whole thing sounds a bit like sloppy language used to create marketing hyperbole… To me, AI is about unsupervised learning.

AI

Manufacturers.

AI = artificial intelligence

Consumers.

AI = artificial information

One day in the not to distant future AI will really “think for itself”

Oticon has been using DNN since they released the More in 2020. It had 12 million sound scenes on the Polaris chip.

The Intent supposedly doubled it on the Intent Sirius platform.

I may be wrong, but isn’t Phonak Autosense AI based?

Right. So the “deep/unsupervised machine learning” process (AI) is what happened at the HA developing stage. The situation seems comparable with the pioneering AI that produced the chess AlphaZero program. Traditionally people would code a chess program using well-established chess guidelines and principles (and some people would considered such clever machines “AI” already back in the 1990s, constantly evaluating positions that no human had yet seen). AlphaZero instead just played zillions of games against itself and self-programmed. That was done on a mainframe, but the resulting program runs, in the end, on very modest hardware- for it no longer needs to run all those test games in real-time.
So- and this may come down to semantics, at some extent- it still seems to me that HAs do not actually (actively) use AI in the modern sense, even though they are smartly programmed. At a minimum, that would require them to self-reprogram their software in real-time, based on listener feedback.

That’s a good analogy!

It sounds like you’re asking for continual learning, AI that keeps learning from your usage. In fact, not all well known LLMs are technically continual learning. Gemini and Grok are to certain extent, but ChatGPT and Mistral are not.

ChatGPT is based on static models with conversational adaptivity, i.e. continually adjusting its prompts or repeating previous prompts behind the scene, rather than continually adjusting neural weights in its model. Simply speaking, it forgets how it answered. You need to keep reminding it what you two have spoken (user interface quietly does that for you) and tell it more to get better answers. It just doesn’t learn on the fly.

AI aids today focus on adaptive behavior (e.g., scene detection, noise reduction, beamforming) rather than on-device continual learning. Power, latency, and privacy constraints make real-time training tough. Personally, I’m hopeful we’ll see personalized, on-device models in hearing aids – per-user profiles and federated learning, so the device gets better for the user over time. We’re getting closer, but not there yet.

Indeed, what machines are doing today forces us to sharpen our definition of “intelligence”. To me, intelligence is still defined by self-learning- something we, and quite a few animal species are capable of. And, yes, big computers, too, as some are now capable of making themselves even smarter than humans for some tasks, using unsupervised “training”. The examples I know of are crystallizing huge datasets (like a snaphot of all of the internet) into simple answers to users on-the-fly, and image processing (e.g. picking out malignant breast tissue histology). And, I guess, building speech audio databases that HAs can use.
However, I would not call a, say, AI-designed mousetrap “intelligent”, in the sense that a mechanical device can be smarter than, for example, an ant. And HAs do not (yet) carry “intelligent chips”. I may be biased as a biologist, and some here (and marketers!) may disagree and still claim that new HAs “run on AI”.
The debate gets even murkier when we ask if a machine is conscious… but let’s not go there…

Image processing…thats a CNN in Neural Network parlance. This is doing exactly the same but its sound processing and deeper, ergo it must be intelligence too

I think its a lot more than an intelligent mousetrap and agree on conscious

There is definitely a dedicated DNN chip in these hearing aids, they are doing more computation than the image processor you mentioned above…

If the definition of intelligence is getting better at doing a task from practice / training from a real world dataset then they are of course intelligent, as are humans, some of those more than others…

Perhaps Smart would be a better description, if a car can drive is that intelligence? Or is it a Smart Car? Is Smartness a form of intelligence?

We can go deep on philosophy I am enjoying this debate.

CNN vs DNN for visual understanding of images3

CNN (Convolution Neural Networks)

Personally I think one of the better definition of Intelligence is still the one put forward by Turing

From wikipedia…

The Turing test, originally called the imitation game by Alan Turing in 1949,^[2] is a test of a machine’s ability to exhibit intelligent behaviour equivalent to that of a human. In the test, a human evaluator judges a text transcript of a natural-languageconversation between a human and a machine.

Originally IBM defined this as a machine being able to beat a Chess Grandmaster, it was one of the first attempts at real AI, again from Wikipedia

Deep Blue was^[a] a customized IBM RS/6000 SPsupercomputer for chess-playing. It was the first computer to win a game, and the first to win a match, against a reigning world champion under regular time controls. Development began in 1985 at Carnegie Mellon Universityunder the name ChipTest. It then moved to IBM, where it was first renamed Deep Thought1, then again in 1989 to Deep Blue. It first played world champion Garry Kasparovin a six-game match in 1996, where it won one, drew two, and lost three games. It was upgraded in 1997, and in a six-game re-match it defeated Kasparov by winning two games and drawing three. Deep Blue’s victory is considered a milestone in the history of artificial intelligence1and has been the subject of several books and films.

So a hearing aid that can make speech in noise better by learning from applying a dataset from a real world environment that can be upgraded by being fed more examples… intelligence? I say yes.

Need to get on with life now but thank you Rob, I always love your posts