How Modern Hearing Devices Separate Voices From Background Noise

Picture a table for four in a busy restaurant. Cutlery rings on plates, a coffee machine hisses, and at the next table someone is telling a story with great enthusiasm. Your companion leans in to speak. A healthy ear and brain will, more or less effortlessly, lift her voice out of the din. For many people with hearing loss, that is exactly the moment the conversation slips away.

Hearing aids have tried to help for decades. Two recent arXiv papers show where their toolkit is heading: one describes a noise-reduction network that first listens to the room, the other surveys how to judge whether processed speech is understandable without asking a human.

Two microphones and a delay

The oldest trick is also one of the most effective. Many modern hearing aids carry two microphones, one slightly in front of the other. As Laurel Christensen explained in The Hearing Review, the processor delays the rear microphone's signal and subtracts it from the front one. Sound arriving from behind reaches both microphones in a way that largely cancels; sound from in front survives. The result is a hearing aid that "points" forwards.

In a test chamber, hearing-aid directional systems typically achieve a directivity index of 2 to 6 dB. In real rooms the benefit is more modest: roughly a 3 dB improvement in the signal-to-noise ratio for sounds from the front, sometimes more when most of the noise is behind you.

Directionality has limits. It works best when the talker is in front, within about 10 feet, and spatially separate from the noise. It can also leave the wearer feeling as if they have tunnel vision for sound. Modern "adaptive" systems steer their point of greatest cancellation towards the dominant noise and switch modes automatically, though in diffuse chatter they behave much like a fixed pattern.

Many channels, many volume knobs

The second tool is compression, and it solves a different problem. Speech spans a wide range of levels, from whispered consonants to booming vowels. Many people with sensorineural hearing loss have a reduced dynamic range: a narrow gap between the quietest sound they can hear and the loudest they can tolerate. Amplify everything equally and soft sounds become audible while loud ones become painful.

As Pamela Souza set out in a classic review in Trends in Amplification, compression turns the gain down as sounds get louder. A 2:1 ratio, for instance, means a 2 dB rise in input produces only a 1 dB rise in output. Multi-channel compression splits sound into frequency bands and compresses each independently, so a fitting can follow an uneven audiogram, giving more help at the pitches where hearing is worst.

It is not a free lunch. Very fast compression can produce a "pumping" effect and blur the rhythm of speech. Many channels can flatten the peaks and valleys of the spectrum that help us tell vowels apart. And Souza's review notes that in background noise, compression generally offers little advantage over simple amplification. Making sounds audible is not the same as separating them.

A filter that learns

Separating them is where deep learning comes in. One influential design is DeepFilterNet, developed by Hendrik Schröter and colleagues at Friedrich-Alexander-Universität Erlangen-Nürnberg and presented at ICASSP 2022. It works in two stages. First, it estimates gains on the ERB scale, a frequency scale modelled on human hearing, to reshape the broad envelope of the sound. Second, it applies "deep filtering" to the lower frequencies, where most of the periodic energy of voiced speech sits, restoring the fine harmonic detail. Sparse, heavily grouped layers keep it light: conference material puts it at under two million parameters.

But a compact network must handle every acoustic world, from kitchen to motorway, with one fixed set of weights.

A fingerprint of the room

That is the problem tackled by DFingerNet, a paper accepted at ICASSP 2025 by Iosif Tsangko, Björn Schuller and colleagues at the Technical University of Munich and Imperial College London, with Michael Müller and Hendrik Schröter of the hearing-aid maker WS Audiology in Erlangen.

The idea is intuitive: let the network hear the room before it hears the conversation. DFingerNet adds a short "fingerprint" recording of the background noise, taken from the same environment but not the exact noise in the mixture. In training, the fingerprint was simply the first second of each noise sample. A separate encoder turns that recording into a compact summary, which is added to the main network's view of the incoming sound. The authors suggest the fingerprint encoder could run on a paired smartphone or smartwatch and stream its summary to the hearing aid, keeping the extra computation off the ear. That is a proposal they have not yet tested in practice.

On a test set built from VCTK speech and FSD50K noises, DFingerNet improved SI-SDR, a measure of how cleanly the voice is recovered, by 11.35 dB, against 10.84 dB for the original DeepFilterNet. Quality (PESQ) gains rose from 0.31 to 0.39. Fingerprints recorded up to two minutes before the speech still worked with little loss. The authors are frank about the limits: the gains are modest, the tests use synthetic mixtures rather than real hearing-aid recordings, and on the steadier noises of the DEMAND set, the original DeepFilterNet actually did better.

:::note How DFingerNet works

• Listen first: capture a short sample of the room's background noise
• Summarise: a separate encoder turns it into a compact "fingerprint"
• Condition: the fingerprint is added to the main network's internal picture of the sound
• Result: +11.35 dB SI-SDR vs +10.84 dB for DeepFilterNet on a mixed-noise benchmark
• Caveat: tested on synthetic mixtures; steady noise still favoured the original :::

Judging without a listener

How do engineers know any of this helps? Listening tests with real people are the gold standard, but they are slow and costly, so researchers lean on predictors.

The established ones are intrusive: they compare the processed speech with a clean reference. The best known in hearing-aid research is HASPI, the Hearing-Aid Speech Perception Index developed by James Kates and Kathryn Arehart at the University of Colorado. Its second version (2021) passes both signals through a model of the ear, which can include the listener's hearing loss, and compares how the speech envelopes are modulated. The catch is obvious: in a real restaurant there is no clean reference.

Hence the growing field of non-intrusive prediction, surveyed in 2025 by Ryandhimas Zezario of Academia Sinica in Taipei. These systems estimate intelligibility from the processed signal alone, usually alongside the listener's audiogram. Some feed the audiogram in directly as numbers; others use it to drive a hearing-loss simulator such as the MSBG model, which reshapes the audio as an impaired ear might hear it. Many now draw on large pre-trained speech models such as HuBERT, WavLM and Whisper, with Whisper's features, the survey reports, tending to give the richest results.

The testing ground has been the Clarity Prediction Challenge, part of a UK research programme. Its first round in 2022, which drew 15 systems, used listening tests in which hearing-aid users repeated what they heard. The second found that non-intrusive systems built on large pre-trained acoustic models could perform well. The third, reported in Rotterdam in August 2025, moved into fully dynamic scenes with real backgrounds and measured hearing-aid signals. Twenty-one submissions came from 14 teams, and the winner predicted the percentage of words understood with an RMSE of 24.98 and a correlation of 0.80.

“

In a real restaurant there is no clean reference.

That is good, not perfect. Zezario's survey lists the hurdles: generalising to new listeners, rooms and devices; the cost of large speech models; and training goals that better reflect what people actually perceive.

Sources & further reading

• Tsangko et al., "DFingerNet: Noise-Adaptive Speech Enhancement for Hearing Aids", arXiv:2501.10525
• Zezario, "Non-Intrusive Intelligibility Prediction for Hearing Aids: Recent Advances, Trends, and Challenges", arXiv:2509.03017
• Schröter et al., "DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering", arXiv:2110.05588
• Barker et al., "The 3rd Clarity Prediction Challenge", ISCA Clarity 2025
• Christensen, "The Evolution of Directionality", The Hearing Review (2013)
• Souza, "Effects of Compression on Speech Acoustics, Intelligibility, and Sound Quality", Trends in Amplification (2002)
• Kates and Arehart's HASPI publications, University of Colorado

✦

Filed under Hearing & Sound. Spotted an error? Tell the editors.