Rewarding a Voice Does Not Make It Easier to Hear in a Crowd

A voice linked to rewards might seem destined to win the brain’s attention. If a person repeatedly earned points whenever a particular speaker talked, would that speaker’s voice become easier to follow the next time several people spoke at once? A new study suggests the answer is usually no. Across three experiments, listeners generally did not understand speech more accurately when it came from a voice associated with higher rewards, nor did a high-value voice become a more distracting masker when listeners were instructed to ignore it. The results challenge the idea that the brain’s powerful reward-learning machinery automatically gives priority to complex sounds such as human voices.

The study, published by Sung-Joo Lim and Lingyu Zi of Binghamton University in Attention, Perception, & Psychophysics, examined a classic auditory challenge known as the cocktail-party problem. In a busy room, the ears receive overlapping streams from multiple speakers, yet listeners can often select one conversation and suppress the rest. This requires the brain to separate the incoming mixture into distinct auditory objects, identify the acoustic characteristics of each speaker, and allocate limited attentional resources to the chosen stream. Previous research has shown that familiar voices can be easier to understand amid competing speech. The researchers asked whether learned value could create a similar advantage—even when the voice itself was not personally familiar.

The idea was grounded in the theory of value-driven attention. In vision, objects associated with monetary rewards or points can later attract attention automatically, sometimes even when they are irrelevant to the task. Similar effects have been reported in hearing: tones and individual words previously paired with large rewards can interfere with the processing of other sounds. Reward is thought to alter the priority assigned to sensory representations, potentially through learning-related changes involving dopaminergic systems and attention networks. But a voice is not a simple tone or isolated visual feature. It is a dynamic, multidimensional acoustic pattern containing pitch, formant structure, speaking rate, intonation and language-specific information. The new research tested whether reward associations could make such a complex auditory object more salient.

In the first experiment, 25 young adults with self-reported normal hearing listened to recordings of four female speakers. Over 1,280 trials spread across two sessions, participants identified which speaker had produced short sequences of digits. Two voices were designated “high value” and two “low value.” Correct identifications of high-value voices produced 10 points on 80 percent of trials and two points on the remaining 20 percent. The probabilities were reversed for low-value voices. The points were displayed on screen but had no monetary value. Because every voice was presented equally often, the researchers attempted to separate reward from simple exposure or familiarity.

After training, participants heard two five-word sentences simultaneously, each spoken by a different voice. One sentence was the target and the other the masker. A visual cue indicated whether the target came from the left or right, and listeners selected the five words they believed belonged to that sentence. The researchers varied the target-to-masker ratio, or TMR, by changing the target’s volume while keeping the masker at 70 decibels sound pressure level. A TMR of –20 decibels meant that the target was much quieter than the competing speech, whereas –5 decibels represented a less extreme imbalance. The streams were also given modest spatial separation through interaural timing differences, simulating voices arriving from locations roughly 30 degrees apart.

The physical difficulty of the listening task had a clear effect: as the target became quieter relative to the masker, listeners identified fewer words. But reward history did not reliably change performance. High-value target voices were no easier to understand than low-value target voices, and high-value masker voices did not produce more confusion. During the training itself, participants were not significantly more accurate or faster when identifying high-reward speakers. Their four-choice responses also showed no bias toward selecting high-value voices. Explicitly learning that some voices were worth more points, therefore, did not appear to make those voices more perceptually powerful in a later speech-comprehension task.

The researchers then changed the way reward associations were learned. In Experiment 2, a new group of 22 participants performed an auditory two-back working-memory task. They heard sequences of three digits and judged whether the middle digit matched the middle digit from two trials earlier. The task required attention to speech content, while the speaker’s identity was irrelevant. Feedback nevertheless paired correct responses with the same probabilistic high- and low-reward values assigned to the four voices. Each speaker was presented in separate blocks, reducing the constant switching between voices that can make speech processing more difficult. Importantly, none of the participants reported noticing that particular voices were associated with particular reward levels.

This incidental, unconscious learning produced the study’s one notable reward-related effect. In the most difficult spatially separated condition, where the target was 20 decibels quieter than the masker, participants were significantly more accurate when the target voice had been associated with high rewards than when it had been associated with low rewards. The advantage disappeared at the three less adverse TMR levels. Reward still did not make high-value masker voices more disruptive, however. The finding suggested that implicit experience might enhance a target voice’s speech representation under severe listening pressure, but only modestly and only in a narrow set of circumstances. The researchers also found that the benefit was significantly larger after incidental training than after explicit voice-identification training.

Experiment 3 tested whether the fragile advantage would become stronger in an even harsher acoustic environment. Thirty new participants underwent the same implicit two-back training, but the target and masker voices were then presented from exactly the same spatial location. Removing the interaural timing difference eliminated an important cue that normally helps the brain separate competing streams. Because this manipulation increases informational masking, the researchers shifted the TMR range upward, from –10 to +5 decibels, to avoid making performance uniformly impossible. The two same-gender voices were otherwise drawn from the same speech materials, and participants used the first word of a sentence to identify which stream to report.

The result was decisive: the reward-associated voices no longer affected comprehension. Neither high-value target voices nor high-value masker voices produced reliable benefits or costs at any TMR. Instead, performance was governed by the acoustic relationship between the streams. Curiously, accuracy was not simply highest when the target was louder. It was lowest when the two voices had equal intensity, at 0 decibels TMR, and highest at –10 decibels, where the intensity difference was greatest. Masker-confusion errors followed the same pattern, peaking when the streams were equally loud. This non-monotonic result indicates that perceptual similarity between two same-gender voices can be more important than target volume alone: when both streams sound acoustically similar and arrive from the same place, the brain struggles to keep them apart.

Taken together, the three experiments provide little evidence that reward attached to a speaker’s identity can reliably override the acoustic constraints of speech perception. The researchers argue that value-driven attention may work most effectively when rewards are paired with simple, stable and highly discriminable stimuli, such as colored visual targets, pure tones or particular words. A human voice changes continuously across syllables and sentences, and the features that identify a speaker do not always overlap with the features that support understanding speech. Training listeners to recognize who is speaking may cause them to emphasize fundamental frequency, vocal-tract resonances or formant patterns without necessarily improving their ability to decode the words.

The findings do not mean that learning is irrelevant to hearing. Familiarity with a spouse’s or friend’s voice can improve speech intelligibility, and short-term exposure can sometimes produce measurable benefits. But familiarity may build detailed talker-specific acoustic and phonetic representations in a way that reward points alone do not. The study also highlights the importance of spatial hearing: when competing voices are separated in space, listeners can exploit binaural cues to reduce masking; when those cues vanish, the auditory system appears to depend heavily on intensity differences and other physical distinctions. Reward, even when learned implicitly across more than a thousand trials and two days, was not strong enough to replace those cues.

The authors caution that their experiments were not preregistered and involved relatively young university participants, so the results should not be interpreted as the final word on motivation and speech. Future studies could test personally meaningful rewards, emotionally significant voices, longer training, different languages, older listeners or people with hearing loss. Researchers might also manipulate the acoustic distinctiveness between speakers to discover when learned value begins to matter. For now, the message from the virtual cocktail party is strikingly clear: a voice may be worth ten points, but when two people speak over each other, the brain still relies primarily on where the sounds come from, how loud they are and how acoustically separable they remain.