Hearing Aids Are Getting On-Device AI — and the Hard Part Is Running It Locally

October 14 — A Counterpoint report, “Global On-Device AI Wearable Forecast,” says global sales of prescription hearing aids are expected to pass 31 million units by 2030, a compound annual growth rate of 4% from 2025 to 2030. As AI spreads through the category, nearly 50% of hearing aids are expected to carry on-device AI by 2030.

For years hearing aids have improved in size, connectivity and battery life, yet the core problem remains: users struggle to follow conversation in noisy places. Traditional hearing aids rely mainly on digital signal processing (DSP): microphones pick up the surrounding sound, a processor analyses features such as frequency and volume, and preset algorithms decide which sounds to amplify or reduce. That works for steady noise such as air conditioning or an engine, but real situations are more complex. In a crowded restaurant, several conversations, music and clinking cutlery occur at once, and rule-based systems struggle to work out which sound the user actually wants.

AI offers a different solution: instead of relying on preset rules, AI models can learn to distinguish the patterns of human voices from other sounds. A deep neural network (DNN) trained on millions of samples of speech and background noise can identify a speaker’s voice in real time and suppress irrelevant noise.

AI can improve a hearing aid’s ability to pick out speech in real environments by separating voices from noise, prioritising the target speaker, improving spatial awareness and adapting to changing surroundings. AI is especially good at telling apart subtle, easily confused sounds such as “s”, “f”, “t”, “sh” and “th”, and at helping users distinguish numbers such as 15 and 50, or 17 and 70 (note: “fifteen” and “fifty” sound similar in English). By preserving these fine details of speech, AI does not merely make things louder — it improves real understanding in restaurants, meetings and crowded streets.

In clinical “speech in noise” tests, listeners with normal hearing usually need a signal-to-noise ratio (SNR) of about 2 dB to identify 50% of key words correctly; for listeners with mild to moderate sensorineural hearing loss, that requirement rises to about 8 dB, and AI can deliver a measurable improvement. According to tests by Phonak, fixed directional microphones with conventional settings improve SNR by 6.4 dB compared with omnidirectional microphones; adding AI noise reduction brings the gain to 9 dB in the default setting and up to 10.2 dB at maximum — enough to converse comfortably in a restaurant.

AI also makes hearing aids more automatic and personal: machine learning adapts sound processing to a user’s preferences and listening situation, cutting manual adjustments and the mental effort of following a conversation, which reduces listening fatigue and how often users ask people to repeat themselves. These gains are not free: running a neural network continuously drains the battery and takes up space, so the most capable AI hearing aids today come in only a handful of styles.

Hearing aids are extremely sensitive to processing delay: the whole path, from picking up sound to delivering it to the ear, must stay within 10 milliseconds for a natural experience. That sets them apart from most AI devices — a smartphone can send data to the cloud and wait for a result, but a hearing aid cannot afford a round trip; even a tiny delay makes what users see and hear fall out of sync, or makes their own voice sound unnatural.

Real-time speech processing therefore requires the AI to run inside the hearing aid rather than on the user’s phone or in the cloud. That is the core value of on-device AI: sound is analysed close to the microphones and receivers.

The difficulty is that hearing aids have tiny batteries, limited computing power and strict size limits. Fitting AI into such a device requires highly optimised models and processors to separate speech and cut noise while keeping power consumption as low as possible. Running a neural network continuously shortens battery life sharply: Phonak’s 2025 Audéo Infinio Ultra Sphere lasts 56 hours in normal use but drops to just 10 hours with AI features running continuously — a vivid illustration of the cost of local AI.

On-device AI makes hearing aids “smart”, but integrating it efficiently is the hard part: every gain in speech clarity costs battery life and space, as Phonak’s figures show. The next generation will compete on how well the AI can run quietly in the background — clearer speech and less listening effort without draining too much power or making the device bigger. The industry has already begun that race.