In 1953 Colin Cherry asked what became known as the cocktail party problem: in a room full of simultaneous talkers, how does a listener follow just one of them? (Cherry, 1953) The question sounds psychological. At bottom it is a problem about information. Sounds are separate at their sources, but by the time they reach a microphone or an ear they have collapsed into a single mixture. A human gets two slightly different mixtures, one per ear, and the auditory system exploits that difference. A single-microphone machine usually gets one. Either way the mathematics is unforgiving: recovering sources from a mixture is ill-posed, because infinitely many combinations of sources could have produced the same signal. Without extra information, there is no way to pick the right one.
The brain manufactures that extra information constantly. Interaural timing and level cues are computed first not in the cortex but in brainstem nuclei such as the medial superior olive (Grothe et al.), then combined with voice timbre, pitch, the predictability of language, and attention itself. Out of the mixture, one voice is pulled forward. Machines never had an equivalent of this, and the gap shows up in three places.
It shows up in the machines themselves. Speech recognizers have reported accuracy approaching human transcription on certain controlled benchmarks (Xiong et al., 2016) — a claim that holds inside those datasets and protocols, and degrades sharply the moment talkers overlap. Voice assistants, meeting software, call analytics, field recordings: nearly everything built on speech assumes one clean speaker, and the real world rarely cooperates. Until the mixture problem is handled, the whole chain is broken at its first link.
It shows up in anyone who has to communicate in noise. The waiter taking an order over the din, the pilot picking a callsign out of radio chatter, the physician listening to a patient in a crowded ward. Selective listening in noise is tiring and error-prone even for people with perfect hearing, and every application that records, transcribes, or translates conversation inherits the same fragility.
And it shows up, hardest of all, in people whose hearing is compromised. The WHO counts more than 1.5 billion people living with some degree of hearing loss (WHO). Modern hearing aids help — directional microphones and noise reduction genuinely improve the signal — but they do not reliably separate overlapping talkers or pick out the one you mean to hear (evidence review). A further group struggles in noise despite normal audiograms; synaptic damage is one debated explanation among several, and the evidence is mixed (Guest et al., 2018). Whatever the cause, the outcome is the same: conversation in a crowd becomes hard to follow, and the cost is not comfort. Epidemiology links hearing difficulty to social isolation and cognitive decline; the 2024 Lancet Commission lists hearing loss among the largest modifiable dementia risk factors. That is an association, not a promise — no hearing technology can claim to prevent dementia — but it says something about the stakes.
So the problem, stated plainly: selective hearing is a basic capability of human perception that machines lack. Sonacern is not a device. It is an attempt to build that capability itself — a core AI that, given a mixture and a cue for whom to listen to, extracts that source. Anything built on speech in the real world can stand on such a core: hearing support, meeting tools, field recording, translation pipelines.
An ill-posed problem yields only to extra information. Sonacern borrows its extra information from the familiarity cue in human perception: give the system a few seconds of the target speaker's voice. The literature calls this target speaker extraction (survey: IEEE Signal Processing Magazine, 2023), and the framing dissolves the two classic dead-ends of blind source separation in one move. You no longer need to know how many talkers are in the mixture, and you no longer need to guess which separated channel was the one you wanted.
Conceptually there are three parts. An encoder turns the reference sample into a compact representation of the speaker — trained to be as robust as possible to what is being said, though in practice phonetic content, microphone and room character do leak in; that is a fact the design fights, not a feature. A separation network processes the mixture, in time-frequency or directly on the waveform. A conditioning mechanism injects the speaker representation into one or more stages of that network — single adaptation layers, multiple blocks and attention mechanisms all appear across system families. The shared principle: steer the network to amplify energy matching the signature and suppress the rest.
What gives such a system scientific worth is not how it is described but how it is tested. We separate four layers of evidence.
Signal evidence. Similarity of the extracted audio to the target, reported with field-standard distortion and intelligibility metrics and with confidence intervals. Cherry-picked audio demos are not evidence.
Task evidence. Feed the extracted audio to a speech recognizer: word error rate should drop, significantly, against both the raw mixture and competing methods. Any number here earns the status of a scientific claim only together with its dataset, noise level, baseline and statistics.
Selectivity evidence. The easiest layer to skip, and the most important. Is the system actually using the reference, or acting as a generic denoiser? Two experiments answer this. Hand it a reference belonging to a different speaker who is present in the mixture: a correct system extracts that speaker instead — proof the output is steerable. Then hand it a reference belonging to someone absent from the mixture: a correct system suppresses its output. It goes quiet. The absent case needs its own metrics, since standard signal ratios degenerate when the true target is silence (Borsdorf et al., 2021).
Human evidence. Objective metrics do not always move together with human perception. Claims aimed at end users need controlled listening tests — with hearing-impaired participants if that is the audience.
None of the above is aspiration; it is the process we ran. The current prototype generation — a reference-conditioned network operating in the time domain — was evaluated end to end on a controlled synthetic testbed: one hundred items, two-talker mixtures with noise, speakers unseen in training. Three of the four evidence layers are on record. The fourth, human listening tests, has not been run yet and sits at the top of the backlog. The numbers below come from our internal experiments; raw logs and the decision journal are archived, they have not passed peer review, and they are open to independent verification against those records.
The same process recorded the failures with equal clarity, and they are the roadmap. Performance drops by 8–12 dB in synthetic reverberation, and real-room generalization is unmeasured. With three simultaneous talkers the system falls short at the hardest noise level. When the target speaker is absent, the system amplifies instead of going quiet; a first targeted training round improved median suppression by more than 50 dB yet still missed the success criteria we had fixed in advance — so it was recorded as "phase two designed," not as "solved." Noisy reference samples and real-time on-device operation remain open fronts. Each is narrowable. None will be called closed without the experiment that closes it.
Three reasons, none of them tied to a single product.
The technological one. Speech keeps moving toward the center of human-machine interaction, and the whole movement leans on an assumption of clean input. A selective-hearing layer removes the assumption. Transcription, translation, archiving, voice control, communication aids — all of them become noise-robust on top of the same core. Solve the mixture problem once, properly, and the gain propagates through the entire chain.
The human one. Failing to follow speech in noise means gradual exclusion from conversation. The heaviest burden falls on people with hearing difficulty, but the tired waiter and the physician in a loud ward live milder versions of the same problem. Given the epidemiological weight behind hearing and isolation, staying in the conversation is, by itself, an outcome worth defending — no exaggeration required.
The scientific one. Going from mixture to source is one of perception's fundamental questions. Every rigorous engineering answer — which cues suffice, how selectivity is proven, when a system should stay silent — feeds back into the understanding of hearing itself. The traffic between auditory neuroscience and extraction architectures now runs in both directions; attention strengthening cortical representations of an attended talker has a striking parallel in networks conditioned on a reference.
A final word on method. This is a rare field where claims can be checked with numbers: what a system does, in which mixture, by which metric, within which confidence interval — all of it definable and reproducible. Sonacern's commitment is that discipline itself. No number spoken before it is measured; every measured number standing next to its context and its raw record. Machines can be taught selective hearing. The way there is to put the proof before the claim.