Sleep has long been recognized as a critical partner in memory consolidation, the dynamic process through which newly acquired information is progressively stabilized and reorganized. During non-rapid eye movement (NREM) sleep, two hallmark brain events dominate the landscape: slow oscillations (SOs), large, slow waves that sweep across the cortex roughly once per second, and sleep spindles, brief bursts of rhythmic brain activity that last about one to three seconds. Prior work has consistently linked memory performance to the power and density of these events, and more specifically to the precise timing with which spindles nest within the SO cycle. Fast spindles, those in the 12 to 15 Hz range, tend to couple with the up-state of the SO, while slow spindles, in the 9 to 12 Hz range, tend to align with the down-state. This cross-frequency coupling is increasingly viewed as the mechanistically specific marker of sleep-dependent consolidation, more informative than the density of either event alone.

Yet the extent to which these neural markers relate to changes in speech-perceptual learning over time has remained under-explored. Speech-perceptual learning is thought to engage both declarative and procedural memory systems, leaving open questions about how sleep contributes to changes in speech perception. Some studies show that talker-specific perceptual learning remains stable over 24 hours, while others report improvements following daytime rest independent of sleep. In synthetic speech learning paradigms, sleep has been shown to restore performance declines observed during wake and protect newly learned mappings from forgetting. For nonnative speech contrasts such as the Hindi dental-retroflex distinction, overnight sleep, but not equivalent wake intervals, has been associated with stabilization against interference and improved identification of novel talkers. Notably, improvements in identification do not always correspond to gains in discrimination, because identification requires a listener to assign a single token to a learned category, whereas discrimination requires only a same-different judgement between two tokens and can in principle be performed on acoustic detail alone. These tasks index qualitatively different perceptual skills and may point to distinct effects of sleep on task performance.

To probe these questions, a team of researchers from the University of Delaware, Georgetown University, and Florida Atlantic University recruited 29 monolingual English-speaking adults aged 18 to 35, spanning a broad range of language ability from those with typical language histories to those with a history of DLD. DLD is a neurodevelopmental disorder that affects language acquisition in approximately 7 percent of children in the United States, and beyond language learning difficulties, individuals with DLD also experience deficits in speech perception that can further impact communication. Participants were classified as having DLD if they reported a history of receiving services and obtained scores that resulted in a positive index for DLD under methods devised by Fidler and colleagues, which utilize raw scores of a 15-word spelling and modified token tests. Participants were assigned to the typically developed group if they indicated no history of speech, language, or reading remediation and obtained a negative index. The researchers excluded individuals with current ADHD, since ADHD medication can alter sleep, and screened for dyslexia to ensure that performance on the DLD classification tasks could not be attributed to a disorder of reading.

The experimental task centered on a notoriously difficult nonnative contrast: the Hindi dental-retroflex distinction, contrasting the retroflex /ɖ/ and dental /d̪/ consonants in the context of minimally pairing words. Participants first completed a baseline discrimination task in which they heard two speech tokens and indicated whether the word onsets belonged to the same or different speech sound category. They were then exposed to a familiarization sequence pairing the auditory tokens with novel visual objects, followed by 200 trials of two-item forced-choice identification training with written feedback after each response. Perceptual ability was measured immediately after training, and again after a delay that included a polysomnography-monitored nap. To take advantage of the natural circadian rhythms of sleep propensity that peak in the early afternoon for this age group, all participants were scheduled from noon to 5 p.m. for a single session, and the laboratory was dimly lit with low-frequency red light to promote sleepiness.

During the nap, participants underwent a lights-out period of 2.5 hours while EEG, EOG, and EMG were recorded continuously at 500 Hz using a 32-channel electrode cap. The researchers used automated sleep scoring software to classify sleep into stage 1, stage 2, slow-wave sleep, and REM. They then identified slow oscillations by low-pass filtering the EEG signal at 3.5 Hz and detecting negative-to-positive zero-crossings within an interval of 0.8 to 2 seconds, requiring the negative peak amplitude to exceed 1.25 times the averaged negative peak. Spindles were detected by band-pass filtering the signal between 9 and 12 Hz for slow spindles and 13 to 16 Hz for fast spindles, computing the root mean square of the filtered signals, and flagging events where the RMS signal remained above 1.5 times the standard deviation for 0.5 to 3 seconds. Slow spindles were quantified over frontal channels F3 and F4, reflecting their predominantly anterior distribution, while fast spindles were quantified over central channels C3 and C4. SO-spindle coupling was estimated using the continuous synchronization index described by Ladenbauer and colleagues, with coupling estimates restricted to stage 2 within NREM sleep because not all participants entered slow-wave sleep during their rest period.

The behavioral results revealed a nuanced picture. Discrimination performance improved significantly from baseline to pre-nap, confirming that the training itself was effective, but there was no significant change in discrimination across the nap interval. In contrast, identification performance improved significantly across the nap interval, with a main effect of Time that was statistically significant. The magnitude of offline changes to perceptual behavior was not modulated by language status, which differs from a previous account that reported a significant difference between TD and DLD adults in perceptual performance following overnight sleep. One possibility is that a nap is too short a duration of sleep for such group-level observations to emerge. Another may be that the study was underpowered in its DLD sample to detect a significant interaction. Indeed, an exploratory paired t-test conducted on identification scores by group suggests that whereas the TD participants improved, the DLD participants did not.

The sleep physiology results were equally revealing. After controlling for baseline and pre-nap discrimination performance, greater slow spindle density significantly predicted better post-nap discrimination performance, whereas greater SO density significantly predicted poorer post-nap performance. Fast spindle density was not significantly associated with post-nap discrimination performance. Neither the association with slow spindle density nor the association with SO density were moderated by language status. The negative correlation between SO density and changes in perceptual discrimination performance is particularly intriguing. NREM sleep characterized by a high density of SOs is thought to support systems-level reorganization, abstraction, and integration. For speech categories, that process may preserve category-level information while reducing access to token-specific acoustic detail. Higher SO density could therefore accompany stronger abstraction but poorer performance on a task requiring fine-grained acoustic comparison. This is consistent with recent observations that slow-wave sleep promotes the preservation of category-level representations at the expense of item-level information.

Perhaps the most striking finding emerged from the coupling analyses. For slow spindles, the coupling strength of SO-slow spindle coupling was negatively associated with language difficulties in those with DLD, meaning that those with stronger SO-spindle coupling had better language skills within the DLD sample. This correlation, with a Spearman rank-order coefficient of -0.71 and a p-value of 0.047, raises the intriguing possibility that SO-spindle coupling may serve a compensatory or protective role in the context of language disorder. However, given the small number of participants with DLD, this finding should be interpreted cautiously. The researchers also observed a numerically lower slow spindle density in adults with DLD relative to TD, with a p-value of 0.067, and exploratory analyses suggested a potential difference in the scalp topography of fast spindle activity between the two groups. Interestingly, prior reports of spindle activity in populations with learning disabilities have suggested that children with dyslexia have higher sleep spindle density relative to TD children, though these earlier studies did not examine spindle topography.

The study has several important limitations that the authors acknowledge. The DLD group was substantially smaller than the TD group, limiting the ability to conduct meaningful between-group comparisons. Nap sleep may not provide sufficient time for a complete sleep cycle, including deeper stages of NREM sleep such as slow-wave sleep, and because approximately half of the participants did not reach this stage, the effects observed here may not fully reflect those associated with longer sleep periods. The study also did not include a wake control group, nor did it include baseline polysomnographic measures of sleep, meaning that although the post-nap improvements observed are consistent with sleep-dependent consolidation, the researchers cannot rule out the possibility that comparable gains would have occurred in the absence of sleep. Future studies should recruit larger and more balanced samples, examine the effects of a full night of sleep, incorporate measures of generalization alongside assessments of trained-item perception, and include a wake control group to allow for a more definitive assessment of the specific contributions of sleep to memory consolidation. Nevertheless, these findings point to a potential relationship between sleep physiology and perceptual-linguistic representation, and suggest that the timing of sleep spindles relative to slow oscillations may be a promising target for understanding individual differences in language ability.

Subject of Research: Sleep-dependent consolidation of speech-sound memory in adults with and without developmental language disorder

Article Title: Speech-sound memory consolidation over a daytime nap in adults with and without developmental language disorder

Article References: Solbi, A., Thimmugari, R. G., van Zelst, A. L., & Earle, F. S. (2026). Speech-sound memory consolidation over a daytime nap in adults with and without developmental language disorder. Frontiers in Behavioral Neuroscience, 20, Article 1903903. https://doi.org/10.3389/fnbeh.2026.1903903

Image Credits: AI Generated

DOI: 10.3389/fnbeh.2026.1903903

Keywords: developmental language disorder, sleep spindles, slow oscillations, memory consolidation, speech perception, NREM sleep, polysomnography, Hindi dental-retroflex contrast, SO-spindle coupling, perceptual learning, neurodevelopmental disorders, daytime nap

The answer, according to a team spanning South Africa, France and the United States, appears to be yes, at least in part. Kyle Seakgwa of the University of the Western Cape, Sharon Peperkamp and Emmanuel Chemla of the Laboratoire de Sciences Cognitives et Psycholinguistique in Paris, and Derrin Pinto and Donny Vigil of the University of St Thomas in Minnesota, found that people who grew up speaking languages with phonemic clicks, such as Zulu or Xhosa, are better than monolingual English speakers at spotting paralinguistic clicks buried in English speech. The advantage is modest, and it emerges only under specific conditions, but its very existence offers a rare window into how linguistic experience shapes auditory perception in ways that go far beyond vocabulary and grammar.

To understand why this matters, it helps to distinguish the two kinds of clicks that linguists study. Phonemic clicks are full-fledged consonants in a small number of languages, estimated at roughly 0.5 percent of the world’s languages, nearly all of them in Africa. These sounds are acoustically and articulatorily complex: they involve a double place of articulation, with constrictions at both the front and back of the mouth held simultaneously, and they rely on a non-pulmonic airstream mechanism in which the tongue creates a vacuum and the release of the front constriction produces a characteristic popping or suction noise. Anyone who has listened to Xhosa or Zulu knows how salient these consonants are; they leap out of the speech stream even to untrained foreign ears.

Paralinguistic clicks are a different beast. They occur in many, perhaps all, languages, but they are not part of any phonemic inventory. Phonetically, they lie outside ordinary consonant systems; grammatically, they are not integrated into morphology or syntax; and semantically, they carry a narrow range of meanings, often tied to emotion or discourse management. A dental click in English can signal disapproval, while a click at the start of an utterance can mark that the speaker is searching for a word or launching a new sequence of talk. Researchers have documented such clicks in English, Spanish, French, German, Hebrew, Russian, Swedish and Mandarin Chinese, typically articulated with the tongue tip against the alveolar ridge or teeth, with a simultaneous velar closure. Less common variants include bilabial clicks, similar to blowing a kiss, and lateral clicks resembling the sound some cultures use to urge on a horse.

Despite their ubiquity, paralinguistic clicks are astonishingly hard to detect consciously. In a 2020 study, Vigil and Pinto showed native English speakers pairs of short video clips of semi-spontaneous speech, one containing a search click and one in which the click had been digitally silenced. Participants were asked whether they noticed any difference in the speaker’s oral expression and, if so, to report the timestamp. They succeeded in fewer than 20 percent of cases, a figure the authors argued was below chance. The result was puzzling because paralinguistic clicks are acoustically prominent: they typically occur during pauses and stand out against low-salience fillers like “um.” If listeners cannot hear such a conspicuous sound, something interesting must be happening in the perceptual system.

The new study set out to test whether speakers of click-using languages, which the researchers abbreviate as CU languages, show heightened sensitivity to these sounds. The team administered the same click detection task online, using the identical stimuli and procedure as the original experiment. Eight video clips of nine to ten seconds each, drawn from a scientific presentation by a native American English speaker, were paired so that one clip contained a paralinguistic search click and the other had the click silenced. Four additional distractor pairs contained no clicks at all. Participants watched each pair, separated by a five-second interval, indicated whether they perceived a difference, and if so reported the timestamp of that difference. Crucially, the click removal happened only in the audio track, a detail participants were deliberately not told.

The first experiment recruited 118 monolingual English speakers from the United States through the online platform Prolific, with the design and analyses pre-registered in advance. The researchers refined the original study’s statistical approach in two ways. First, they adopted a more stringent chance level of about 17 percent, reasoning that a participant has roughly a one-in-two chance of claiming any difference exists and approximately a one-in-three chance of landing the timestamp within one second of the click in a nine-to-ten-second clip. Second, they used bootstrapping with 1,000 resamples to estimate 95 percent confidence intervals, allowing a principled assessment of how plausible the null hypothesis really is. A detection was counted as timely only if the participant both correctly reported a difference and placed it within one second of the click.

The English speakers’ performance was strikingly poor. Mean timely detection on target items was just 11 percent, even lower than the 17.4 percent reported in the original classroom-based study, and statistically indistinguishable from the stringent chance level in both the opening condition, where clicks began a sentence, and the mid condition, where they appeared mid-sentence. Accuracy was also lower than in the original study for sentence-initial clicks, which the authors suggest may reflect the challenges of online testing compared with a controlled classroom setting. Yet the picture was not entirely bleak. The confidence interval for timely detections did not include zero, participants reported more differences in target pairs than in distractor pairs, and they described hearing the presence or absence of a click-like sound only in response to target pairs. The researchers conclude that English speakers’ capacity to detect paralinguistic clicks is weak, but not null.

The second experiment turned to South Africa, where Prolific offers no direct filtering by click language, so participants were initially selected as South-born and resident English-speaking multilinguals. Of 124 people tested, 54 reported speaking at least one CU language and were included in the analysis: 46 spoke Zulu, 16 Xhosa, 14 Sotho, 9 Swati and 3 Ndebele, with 27 speaking two or more of these languages. Using the same stimuli and procedure, these participants again showed overall low detection rates, performing no better than chance on sentence-medial clicks. But on sentence-initial clicks, their performance was significantly better than that of the monolingual English speakers, and their responses showed a clearer separation between target and distractor pairs, including click mentions that were entirely absent from distractor trials.

Why would the advantage be confined to clicks at the beginning of sentences? The researchers offer a tantalizing hypothesis rooted in phonology: click consonants in CU languages tend to appear at the beginning of words, so speakers of these languages may have tuned their attention to the positions where clicks are most likely to occur. The broader implication is that the presence of phonemic clicks in a person’s native language yields a perceptual advantage for non-linguistic clicks, at least when listening to a language the person speaks as a second language. The effect is small, and it survived despite an adversarial online setting and short stimuli, which makes its detection all the more notable.

Many questions remain. The team used only search clicks embedded in emotionally neutral speech, leaving open whether stance clicks in emotive-laden speech would be easier to catch. Audiovisual stimuli were chosen to prevent confusion between clicks and mouse clicks, yet audio-only presentation might make detection easier, as one study of the filler “um” suggests. And because all participants listened to English, a language they all knew, the researchers cannot rule out that detection would improve in an unknown language, where clicks are known to be highly salient even to listeners who speak nothing like them. What the study establishes is that the sounds we learn to treat as meaningful consonants reshape how we parse the incidental noises of conversation. In roughly half a percent of the world’s languages, clicks do grammatical work; everywhere else, they whisper at the edges of meaning. This research shows that where clicks live in a language determines, quite literally, what the ear learns to hear.

Subject of Research: Cross-linguistic differences in the auditory detection of paralinguistic click sounds in conversation

Article Title: Cross-linguistic variation in the detection of conversational click sounds

Article References: Seakgwa, K., Peperkamp, S., Pinto, D., Vigil, D., & Chemla, E. (2026). Cross-linguistic variation in the detection of conversational click sounds. Frontiers in Language Sciences, 5, Article 1932401. https://doi.org/10.3389/flang.2026.1932401

Image Credits: AI Generated

DOI: 10.3389/flang.2026.1932401

Keywords: click languages, paralinguistic clicks, phonemic clicks, cross-linguistic perception, Zulu, Xhosa, speech perception, bilingualism, psycholinguistics, phonetics, discourse markers, auditory detection

The study, led by Andrea Pešková of the Institute of Romance Studies at the Free University of Berlin, tested how 104 adult learners of Italian from five linguistic backgrounds identify consonant length, alongside a control group of 21 native Italian listeners. The learners came from Finnish, Czech, Slovak, German and Spanish—languages chosen because they differ dramatically in how they use sound duration. Finnish contrasts both vowel and consonant length phonemically, making quantity a central organizing principle of its sound system. Czech and Slovak encode vowel length but not consonant length, although Slovak additionally allows long syllabic liquids such as /r:/ and /l:/. German has a more restricted system in which vowel duration interacts with vowel quality and is limited to stressed syllables. Spanish has no phonemic segmental length at all, apart from the contrast between the tap and trill in rhotics, as in pero versus perro.

The method was elegantly controlled. Participants listened to 54 trisyllabic Italian pseudowords—nonwords that follow Italian phonotactic rules but resemble no real word in any of the tested languages—recorded by a male native speaker with uniform declarative intonation. Each pseudoword belonged to a minimal pair differing only in the duration of the intervocalic consonant, such as sàpolo versus sàppolo. The stimuli covered nine consonant categories, including voiceless and voiced stops, affricates, and fricatives, as well as sonorants: a lateral, a rhotic and a nasal. Each target consonant was embedded in three prosodic configurations, depending on whether it followed the stressed vowel, preceded it, or sat far from any stress. On each trial, listeners heard a pseudoword and chose between two written options on screen, one spelled with a single consonant letter and one with a double letter, mirroring Italian orthography.

The headline result was a striking accuracy hierarchy that tracked the learners’ native phonologies almost perfectly. Overall identification accuracy across all participants was 81 percent, and native Italian listeners performed at near-ceiling levels of around 96 percent. But the learner groups split sharply: Finnish speakers reached 89.8 percent, Czech 83.3 percent, Slovak 83.1 percent, Spanish 75.1 percent, and German just 64 percent. Every learner group differed significantly from the native controls. The pattern confirms that the first language acts as a perceptual filter—what Trubetzkoy famously called the phonological sieve—through which foreign sounds are interpreted. Learners whose first language already encodes quantity contrasts heard Italian gemination far more reliably than those whose languages do not.

Perhaps the most theoretically interesting finding concerns Czech and Slovak learners. Neither language contrasts consonant length phonemically, yet both groups performed substantially better than German and Spanish speakers. This supports what is known as the Feature Hypothesis, proposed by McAllister and colleagues: transfer between languages can operate at the level of abstract phonological features rather than individual segments. In other words, experience with duration as a linguistically relevant cue—say, in vowels—can generalize to help perceive duration in consonants. This fits earlier work showing that Estonian listeners, whose language has rich quantity distinctions, perceived Swedish vowel length better than English and Spanish listeners, and that vowel-length experience aids discrimination of non-native consonant length.

Yet the Feature Hypothesis does not tell the whole story. The segmental inventory of each learner’s native language also mattered, sometimes in surprising ways. Finnish learners, despite the best overall scores, struggled with the affricates [tʃ] and [dʒ] and the fricative [f], sounds absent from Finnish or limited to loanwords. Spanish learners, though weaker overall, showed a distinct advantage with rhotics, plausibly thanks to their native tap–trill contrast, which itself involves durational differences. Slovak learners did not outperform Czech speakers on liquids as predicted, and both groups showed unexpectedly low accuracy for [f]. Even German learners stumbled on the affricate [tʃ], a sound that exists in German. Quantity experience, the study suggests, is not enough on its own; familiarity with the specific segments carrying that duration matters too.

Consonant type produced another consistent effect. Sonorants—nasals, laterals and rhotics—were identified more accurately than obstruents across all learner groups, with predicted accuracy of 98.1 percent for sonorants versus 96.9 and 96.2 percent for voiced and voiceless obstruents respectively. Sonorants also elicited significantly faster responses. The explanation lies in acoustics: sonorants carry continuous, stable acoustic energy, which makes differences in duration easier to detect, whereas obstruents, with their brief bursts and turbulent noise, provide less robust temporal cues. This universal perceptual tendency interacted with language-specific experience, illustrating that both general auditory factors and native-language structure shape how listeners judge consonant length.

Two factors that might have been expected to matter turned out to play surprisingly small roles. Stress position, which influenced perception in an earlier discrimination experiment with the same participants, showed no significant effect in the identification task for any learner group—only native Italians showed a modest dip in unstressed contexts, and German learners showed a slight post-stress advantage. Proficiency in Italian also had only a modest, non-significant effect, with the influence of the first language dwarfing any gains from advancing from beginner to advanced levels. Even Finnish learners at lower proficiency outperformed most other groups, suggesting that native-language structure provides a powerful head start from the earliest stages of acquisition.

The study also compared identification performance with the same participants’ earlier results on a discrimination task, in which listeners had to judge whether two successive stimuli differed. Identification accuracy was higher across all groups except the Finnish learners. The authors attribute this partly to the different cognitive demands: discrimination taxes auditory memory and fine-grained acoustic processing, while identification taps into phonological categorization and was further aided by written response options. Orthography likely helped as well, since Italian transparently marks geminates with double consonant letters—a cue particularly useful for Finnish learners, whose own orthography also uses double letters for long consonants, and, more subtly, for Spanish speakers familiar with the r versus rr spelling distinction. German orthography, by contrast, uses double consonants to signal a preceding short vowel, which may actually have interfered and contributed to German learners’ weakest-in-class performance.

The broader takeaway is that second-language listeners do not hear foreign sounds through their ears alone; they hear them through their entire linguistic history. Mapping continuous acoustic variation—milliseconds of closure and friction—onto discrete phonological categories is a negotiation between abstract feature experience, specific segmental knowledge, orthographic habits and task demands. For learners and teachers alike, the message is concrete: speakers of quantity languages such as Finnish or Czech have a genuine perceptual edge on Italian geminates, while speakers of German and Spanish may need targeted training on duration as a meaningful cue, ideally through identification-based exercises with written support. Future work from the same project, including production data with and without orthographic cues, should reveal how these perceptual advantages translate into speech—and how much of an accent, ultimately, is written into the language we grew up with.

Subject of Research: Second-language perception of Italian consonant length contrasts and cross-linguistic influence

Article Title: Identification of consonant length in L2 Italian: the role of cross-linguistic influence

Article References: Pešková, A. (2026). Identification of consonant length in L2 Italian: the role of cross-linguistic influence. Frontiers in Language Sciences, 5, Article 1911194. https://doi.org/10.3389/flang.2026.1911194

Image Credits: AI Generated

DOI: 10.3389/flang.2026.1911194

Keywords: Italian geminates, consonant length, second language acquisition, speech perception, cross-linguistic influence, phonological quantity, Finnish, Czech, Slovak, German, Spanish, orthography

The study, published in Frontiers in Psychology, drew on responses from 691 Cantonese-speaking undergraduates in Hong Kong who listened to recordings of speakers reading the same short passage. The listeners rated each speaker’s competence and warmth, the strength of their accent, and how easy they were to understand. Crucially, the researchers also measured how much threat each participant perceived from Mainland Chinese people, a socially and politically salient outgroup in Hong Kong during the period the data were collected, between 2018 and 2020.

The theoretical backdrop is Intergroup Threat Theory, which holds that people who believe an outgroup threatens their group’s resources or values respond with hostility. Threats come in two flavors: realistic threats to status, power, and material resources, and symbolic threats to values, norms, and worldviews. Decades of research show that perceived threat predicts negative attitudes toward outgroups. What makes the new study striking is its suggestion that this hostility leaks into something as seemingly objective as speech perception itself.

The participants each heard eight speakers: some speaking Cantonese, the ingroup language, and some speaking Putonghua, the outgroup’s language. Within each language, speakers were either native or nonnative, with the nonnative speakers producing accented speech shaped by their first language. The design allowed the researchers to test a subtle prediction: threat should bias perception only when outgroup members speak the ingroup’s language, because that is the comparison in which the ingroup can most favorably distinguish itself.

That is largely what the data showed. Using multilevel regression models that accounted for ratings nested within participants, the team found that greater perceived threat predicted lower competence ratings specifically for Putonghua-accented Cantonese speakers – Mainlanders speaking the Hong Kong language. Threat did not meaningfully predict evaluations of native Putonghua speakers, whose superior fluency in their own language was acknowledged regardless of how threatened a listener felt. This selective pattern is important: a general surge of negativity toward outgroup voices would have penalized both, but it did not.

The effects on speech perception were even more revealing. Participants high in perceived threat rated Putonghua-accented Cantonese speech as significantly more accented and significantly harder to understand. They showed no such bias toward native Cantonese speakers, and none toward Cantonese-accented Putonghua speakers. In other words, listeners inflated the perceived speech markers of outgroup members only on the dimension where the ingroup would come out ahead – a maneuver the authors interpret as social creativity, a strategy from social identity theory in which group members redefine comparisons to secure positive distinctiveness.

Mediation analyses then tied the pieces together. The pathway from threat to more negative speaker evaluations ran through biased speech perception: threat predicted stronger perceived accent and lower comprehensibility, which in turn predicted worse competence and warmth ratings. Notably, the comprehensibility leg of this chain was far more consistent than the accent leg. Robustness checks across the three datasets and across different accent levels showed that the accent-strength effect was fragile – appearing only in the 2018 data and only for mildly accented speakers – whereas the comprehensibility effect held steady. Perceived ease of understanding, not accent per se, appears to be the pivotal mediator.

These findings align with the fluency principle in language-attitudes research: speech that is harder to process triggers negative affect and harsher trait judgments. Earlier work has shown that comprehension aids such as subtitles improve evaluations of accented speakers, while barriers like background noise damage them. The new study adds a non-linguistic twist – the listener’s intergroup attitudes can manufacture the very processing difficulty that then justifies their negative judgment. The authors caution that because threat was measured after the listening task and all variables were collected in quick succession, the causal chain cannot be firmly established, and reverse pathways remain plausible.

The practical stakes are considerable, particularly in education and workplaces where accented speakers are evaluated on high-stakes dimensions. If students or instructors are judged through the lens of intergroup threat, assessments of their speech may reflect prejudice as much as performance. The research points toward two promising interventions: increasing exposure to and familiarity with nonnative accents, which reliably improves comprehension and social judgments, and fostering positive intergroup contact, which is known to reduce perceived threat. Bilingual classroom interaction between native Cantonese and Putonghua speakers could, the authors suggest, work through both channels at once.

The study also carries caveats about generalizability. All speakers read the same passage, only two speakers represented each accent level, nonnative accent levels were collapsed in the main analyses, and the data were gathered during a period of extraordinary social upheaval, when perceived threat was likely elevated beyond typical levels. The findings may therefore be a snapshot of a particular moment in Hong Kong’s intergroup history. Still, the core message travels: what we hear is not only a function of what is said, but of who we believe is saying it and how threatened we feel. In linguistically diverse societies, reducing perceived threat and easing the processing of accented speech may be twin keys to fairer listening.

Subject of Research: How perceived intergroup threat biases speech perception and social evaluations of accented speakers in Hong Kong

Article Title: Comprehension under threat: outgroup threat predicts perceived accent strength, ease of comprehension, and speaker evaluations

Article References: Shum, P. L.-C., Tse, C.-S., & Wright, S. C. (2026). Comprehension under threat: outgroup threat predicts perceived accent strength, ease of comprehension, and speaker evaluations. Frontiers in Psychology, 17, Article 1835739. https://doi.org/10.3389/fpsyg.2026.1835739

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1835739

Keywords: intergroup threat, language attitudes, nonnative accents, speech perception, processing fluency, comprehensibility, Hong Kong, Cantonese, Putonghua, social creativity, stereotyping, Comprehension

The study, led by Dr Kazuya Saito-van Poeteren of the UCL Institute of Education, examined 67 Japanese learners of English. The team first measured each participant’s auditory precision by testing how well they could detect slight acoustic differences in non-speech sounds, such as subtle changes in pitch, length and tone. This deliberately simple, domain-general measure of listening skill was then compared against how the participants’ brains responded to foreign speech sounds that their native language does not contain.

To capture the brain activity underlying speech perception, the researchers used functional magnetic resonance imaging, or fMRI. Participants were scanned while trying to distinguish between two English vowels that are notoriously difficult for Japanese speakers: /æ/, the vowel sound in the word cat, and /ʌ/, the vowel in cup. Neither of these sounds maps directly onto any vowel in the Japanese phonological system, which makes them an ideal test case for studying how learners cope with acoustic material that their native language has not prepared them for. In a second task, the same participants distinguished comparable subtle differences in non-speech sounds, providing a direct contrast between how their brains handled foreign speech and how they handled complex acoustic signals stripped of any linguistic meaning.

The results revealed a striking pattern. Learners with higher auditory precision relied less on the brain’s left sensorimotor system, a network heavily involved in language processing, when they listened to foreign-language sounds. Instead, their brain activity patterns resembled those that occur when the brain processes non-speech sounds. In other words, the brains of these skilled listeners did not immediately treat the unfamiliar vowel sounds as language at all. Rather than forcing the new acoustic input into pre-existing linguistic categories, they appeared to analyse it as raw, objective acoustic information. With less interference from the entrenched sound patterns of their native tongue, it became simpler for these learners to process the new sounds accurately.

The contrast with learners who scored lower on auditory precision was equally informative. The fMRI results showed that these participants relied more heavily on the left sensorimotor network when listening to foreign-language sounds. Their brains appeared to filter the unfamiliar acoustic input through the pronunciation patterns of their native language, effectively trying to map sounds like /æ/ and /ʌ/ onto the closest available Japanese categories. According to the researchers, this native-language interference is precisely what makes certain foreign sounds so difficult to perceive and produce accurately, and it helps explain why learners with weaker auditory precision tend to struggle more with pronunciation.

Dr Saito-van Poeteren explained the significance of this mechanism: Learners with a good ear for auditory differences can hear a new foreign-language sound objectively, without immediately processing it as language. Instead, their brains analyse the acoustic information itself, and this may be the key to why some people grasp new languages with more ease than others. The observation reframes a long-standing puzzle in second-language acquisition research. Rather than asking simply whether some learners are better at languages, the study points to a specific, measurable perceptual ability that operates beneath language learning itself.

The implications extend well beyond laboratory curiosity. Most people approach foreign language learning by memorising vocabulary and listening to speech, Dr Saito-van Poeteren noted, but the study indicated that language learners might also benefit from training the auditory abilities that underpin language learning in the first place. On this view, the ear is not merely a passive receiver of linguistic input but an active instrument whose sensitivity can shape how well the brain absorbs new sound systems. A learner whose auditory precision is high arrives at the task of foreign pronunciation with a genuine perceptual advantage, one that conventional study methods neither test nor cultivate.

The research team is already looking toward practical applications. They aim to develop a new form of combined language and auditory training designed to improve learners’ ability to detect fine differences in pitch and other acoustic properties. The idea is that strengthening this domain-general listening skill could, in turn, make foreign speech sounds easier to perceive and reproduce, effectively training the ear before and alongside the language itself. Such an approach would represent a meaningful departure from current practice in many classrooms around the world.

Indeed, Dr Saito-van Poeteren argued that foreign language education should move beyond an approach based simply on more practice and towards one that develops learners’ ear for language learning in line with their individual auditory characteristics. The phrase captures a broader shift in thinking that the study embodies: individual differences in second-language success may be traceable to perceptual and neural profiles that vary from person to person, and teaching could become more effective if it accounted for those profiles rather than treating all learners as though they hear the same acoustic world in the same way.

The study, titled Individual Differences in Motor and Auditory Activation in L2 Speech Learning: Roles of Domain-General Auditory Processing, was conducted by Kazuya Saito-van Poeteren together with Adam Tierney, Kotaro Takizawa, Diego Dardon, Chunlin Liu, Magdalena Kachlicka, Motoaki Sugiura and Hyeonjeong Jeong. It is published in the Journal of Experimental Psychology: General. For anyone who has ever wondered why their accent refuses to budge while a friend seems to slip effortlessly into a new phonetic world, the research offers both an explanation and a hopeful prospect: the difference may lie in a trainable perceptual skill, not in an immutable gift for languages, and the path to better pronunciation may begin not with the mouth but with the ear.

Subject of Research: Auditory precision and its role in second-language speech learning and pronunciation

Article Title: People with a ‘good ear’ for sounds make better language learners

Article References: People with a ‘good ear’ for sounds make better language learners. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: auditory precision, second language acquisition, pronunciation, fMRI, speech perception, sensorimotor system, UCL, Tohoku University, Japanese learners of English, vowel perception, neuroscience, language education

The research, conducted by Zixuan Wu and Carol K. S. To at the University of Hong Kong, focused on a uniquely demanding listening environment: a tone language. In languages such as Cantonese, Mandarin, and Vietnamese, the pitch contour of a syllable carries meaning. Say the Cantonese syllable “ji” with a high, flat pitch and it means “doctor”; say it at a mid level and it means “meaning.” The trouble is that pitch is also the primary acoustic signature of who is talking. Fundamental frequency, the physical correlate of perceived pitch, simultaneously encodes vocabulary and identity, forcing listeners to disentangle linguistic content from personal vocal characteristics within the very same acoustic dimension.

This dual role makes Cantonese level tones a brutal test case for talker adaptation. A high-level tone produced by a low-pitched man can overlap in absolute frequency with a mid-level tone from a high-pitched woman, so accurate perception requires listeners to estimate each talker’s baseline pitch and judge tones relative to that baseline rather than against absolute values. Previous work shows how punishing this can be: when Cantonese listeners heard level tones from seven male talkers mixed together, identification accuracy collapsed from roughly 80 percent to under 49 percent, compared with the mild costs seen when English listeners identify vowels across many voices.

To probe how autistic adults cope with this challenge, the researchers recruited 25 autistic adults and 25 neurotypical controls, all native Cantonese speakers from Hong Kong, matched carefully on age, sex, and nonverbal intelligence. Diagnoses in the autistic group were confirmed with the Autism Diagnostic Observation Schedule, and none of the participants had hearing impairments. Each person completed a speeded word identification task in which they heard the syllable “ji” produced with either a high-level or mid-level tone and pressed a key to indicate which of the two words they heard, responding as quickly and accurately as possible.

The experiment’s elegance lay in its factorial design. The researchers recorded four native talkers, two men and two women, each producing every tone four times at their habitual speaking rate, yielding 32 natural tokens. Across four blocks, participants heard speech from either a single talker or all four, and from either a single exemplar of each tone or all four natural repetitions. This allowed the team to separate two distinct sources of acoustic variability: between-talker variation, the systematic differences tied to vocal identity, and within-talker variation, the small, unsystematic fluctuations that arise even when one person repeats the same word in a neutral voice.

The between-talker results were strikingly symmetric. Both groups showed robust processing costs when tones came from multiple talkers, with lower accuracy, slower responses, and reduced perceptual efficiency, and the magnitude of the cost did not differ between autistic and neurotypical listeners. This finding aligns with the HIPPEA account, a predictive-processing theory of autism proposing that prediction errors are assigned abnormally high and inflexible precision. Under that framework, every talker change generates strong error signals, so autistic listeners remain fully sensitive to systematic acoustic shifts. It sits less comfortably with the competing hypo-priors account, which holds that autistic perception relies less on learned expectations and might therefore be less disrupted when a talker changes.

The within-talker results told a different story. When multiple natural repetitions of the same tones from the same talkers were introduced, only the autistic group showed a significant drop in efficiency; the neurotypical listeners were essentially unbothered. From a Bayesian perspective, this makes evolutionary and computational sense. Small, unsystematic fluctuations within a talker’s productions fall in the high-probability region of that talker’s acoustic distribution and should be down-weighted as noise rather than treated as informative prediction errors. The autistic listeners, however, appeared to remain sensitive to these fine-grained deviations, consistent with the detail-focused cognitive style long described in autism research and with first-person reports from autistic adults who describe noticing very faint sounds that others filter out.

Two further analyses deepened the picture. First, the size of a listener’s between-talker cost did not correlate with the size of their within-talker cost in either group, suggesting the two forms of variability are handled by at least partially dissociable perceptual mechanisms rather than a single gradient process that simply scales with acoustic distance. Second, the comparable between-talker costs in the Cantonese speakers contrast with a 2024 study of English-speaking autistic adults, who showed no multi-talker cost at all. The authors suggest that lifelong experience mapping pitch to both meaning and identity in a tone language may strengthen talker-specific priors, and that the English study’s absence of a cost may reflect differences in automatic attentional orienting to talker change rather than a genuine inability to adapt.

The practical implications reach into everyday life. Socially diverse listening environments, from open-plan offices to family dinners, are saturated with exactly the kind of subtle, unsystematic vocal variation that proved costly for autistic adults in this study. If stabilizing word representations against such fluctuations demands extra perceptual resources, then real-world speech comprehension in noisy, multi-voice settings may be more effortful for autistic listeners than laboratory tests of accuracy alone would suggest. The findings also carry a theoretical punch: perceptual differences in autism may not reflect a generalized weakness in using prior knowledge, but rather a specific difference in how fine-grained sensory deviations are weighted against established perceptual regularities. The authors caution that the sample was modest and limited to Cantonese-speaking adults, and they call for cross-linguistic and developmental studies, including work with autistic children and speakers of other tone languages, to determine how talker adaptation unfolds across different linguistic systems and stages of life.

Subject of Research: Lexical tone perception and talker variability accommodation in autistic adults

Article Title: Challenges in Accommodating Talker Variability in Lexical Tone Perception Among Autistic Adults: Implications for Talker Adaptation

Article References: Wu, Z., & To, C. K. S. (2026). Challenges in Accommodating Talker Variability in Lexical Tone Perception Among Autistic Adults: Implications for Talker Adaptation. Journal of Autism and Developmental Disorders. https://doi.org/10.1007/s10803-026-07531-w

Image Credits: AI Generated

DOI: 10.1007/s10803-026-07531-w

Keywords: autism, speech perception, lexical tone, Cantonese, talker variability, fundamental frequency, predictive processing, Bayesian perception, talker adaptation, auditory processing, language, neurodevelopment

The phenomenon at the heart of the study is known as compensation for coarticulation. Speech sounds are produced in a continuous stream, and neighboring sounds blur into one another. When a fricative such as “s” or “sh” follows a vowel, listeners systematically shift the boundary between the two categories depending on the vowel they heard. After a vowel like “a,” which pushes the tongue body back and lowers the acoustic energy in the high-frequency range, listeners are more likely to categorize an ambiguous fricative as “sh”; after “i” or “u,” the boundary shifts the other way. This context effect, first documented systematically by Mann and Repp in 1980, is one of the most robust findings in speech perception, yet the size of the effect varies considerably from one listener to the next.

The obvious candidate for explaining that variation is personality. In 2010, researcher Alice Yu reported a striking result: neurotypical listeners who scored higher on the Autism Spectrum Quotient, a questionnaire measuring traits associated with the autism spectrum, compensated less for coarticulation. The finding resonated widely, in part because it connected a low-level perceptual effect to broader theories about how autistic traits shape cognitive style, including attention to detail and reduced sensitivity to context. It also fed into a growing literature on individual differences in phonology, suggesting that the seeds of sound change might lie in the perceptual quirks of particular listeners.

Kingston’s study set out to replicate that result with substantially more statistical power. The original work had relied on a modest sample, and the replication crisis in psychology has repeatedly shown that small samples can produce seductive correlations that evaporate under scrutiny. In the new AQ experiment, the number of participants was more than doubled, and the analysis was rebuilt from the ground up using Bayesian multilevel modeling, a framework that estimates full probability distributions over parameter values rather than single point estimates. Listeners completed the Autism Spectrum Quotient, including its subtests covering attention switching, attention to detail, communication, imagination, and social skills, before categorizing stimuli drawn from a synthetic continuum between “s” and “sh” embedded in different vowel contexts.

The outcome was unambiguous. Despite the larger sample and the more sensitive analysis, the specific dependencies that Yu had reported between autistic traits and compensation for coarticulation failed to replicate. Listeners still showed the classic context effects, and their responses still varied from person to person, but the variation did not track their questionnaire scores in the way the original study had suggested. A related study by Lai and colleagues in 2022 had already reported that perceptual compensation and lexical effects were uncorrelated with AQ responses, though a recording problem with their questionnaire had complicated the comparison. The new results strengthen the case that the original correlation was fragile.

But Kingston did not stop at a failed replication. A second experiment, the B5 study, asked a broader question: perhaps the wrong personality instrument had been used. The Big-Five questionnaire, which measures agreeableness, conscientiousness, emotional stability, extraversion, and intellect-imagination, offers a more comprehensive map of personality structure than the AQ, and some researchers have argued that autistic traits are not an independent personality dimension at all but a particular configuration of Big-Five characteristics. If compensation for coarticulation reflects personality in any form, the Big-Five framework should have been able to detect it.

Here the results were more intriguing, and more complicated. Listeners’ categorization responses did differ as a function of their Big-Five trait profiles, and also as a function of their gender. Yet the size and even the direction of compensation varied across genders and across individual traits, with no single, coherent pattern emerging. Rather than revealing a clean mapping between a personality dimension and a perceptual strategy, the data showed a tangle of interactions whose interpretation is far from straightforward. The effects were real in the statistical sense, but they do not add up to a predictive theory of how a given person will hear a given sound.

Kingston draws a deliberately cautious conclusion. The answer to the question posed in the title, he suggests, is no, in one of two senses. Either listeners’ performance is not predictable from their personality traits at all, or it is predictable in principle but the causal pathway remains elusive, with the observed correlations possibly reflecting confounds rather than genuine perceptual consequences of temperament. Distinguishing between these possibilities would require experimental designs that go beyond correlating questionnaire scores with perceptual judgments, perhaps manipulating state rather than trait, or measuring perception repeatedly within individuals across time.

The study also carries a broader methodological message. The original 2010 finding had been cited widely and woven into theoretical accounts of sound change, in which individual perceptual differences were proposed as the raw material on which language evolution acts. If those differences are not reliably tied to measurable personality traits, then theories of sound change may need to find their sources of variation elsewhere, perhaps in the acoustics of different speech communities or in the sheer stochasticity of perceptual processing. The failure also illustrates why large replications matter: a correlation that looks compelling in a small sample can dissolve when the sample doubles and the analysis becomes more rigorous.

In keeping with the open-science practices that have become standard in the field, Kingston has made the data, analysis code, stimuli, and even the script for generating the s-to-sh continuum freely available on the Open Science Framework, allowing any researcher to scrutinize, reuse, or extend the work. Whatever the ultimate verdict on whether perception is personal, the tools for settling the question are now in everyone’s hands. For now, the sounds we hear may differ from person to person, but the reasons for those differences remain stubbornly, fascinatingly opaque.

Subject of Research: Individual differences in speech perception and compensation for coarticulation in relation to personality traits

Article Title: Is perception personal?

Article References: Kingston, J. (2026). Is perception personal?. Attention, Perception, & Psychophysics, 88(7), Article 200. https://doi.org/10.3758/s13414-026-03340-6

Image Credits: AI Generated

DOI: 10.3758/s13414-026-03340-6

Keywords: speech perception, compensation for coarticulation, Autism Spectrum Quotient, Big Five personality, individual differences, replication, Bayesian modeling, psychoacoustics, phonetics, perceptual categorization, sound change, psycholinguistics

The study, led by Seung-Eun Kim of Yonsei University together with Matthew Goldrick, Joseph Keshet, and Ann R. Bradlow of Northwestern University and the Technion-Israel Institute of Technology, tackled a puzzle that has dogged the field for years. When native English listeners adapt to one non-native talker, does that adaptation help them understand a different non-native talker they have never heard before? The literature is inconsistent. Some experiments report robust generalization from a single training voice; others find that the learning stays stubbornly locked to the specific talker. The new work was designed to adjudicate between two competing theoretical accounts, and to do so with an unusually rigorous, data-driven methodology.

The first hypothesis centers on intelligibility. Perhaps, the researchers reasoned, generalization is driven by how easy the training talker is to understand. A highly intelligible talker might provide cleaner, more reliable evidence about how English sounds are realized in non-native speech, allowing the listener’s perceptual system to extract general rules rather than talker-specific quirks. Under this account, training with a clear, easily understood accented speaker should produce stronger benefits for understanding other accented speakers than training with a hard-to-understand one. The second hypothesis centers on acoustic similarity. Perhaps generalization depends on how closely the training talker resembles the test talker in the physical details of their speech. If the perceptual system tunes to specific acoustic properties, then a training voice that sounds similar to a test voice should yield better transfer than a dissimilar pairing. This account echoes earlier findings by Xie and Myers, who showed that acoustic similarity between talkers constrains how far foreign-accent adaptation spreads.

Testing these hypotheses properly required something the field has rarely had: precise, quantitative measurements of both intelligibility and acoustic similarity for a large pool of second-language talkers. The team drew on the ALLSSTAR archive, a Northwestern University repository of recordings from more than 100 speakers of English as a second language, each of whom had been assessed with empirical intelligibility measures. For the similarity dimension, the researchers turned to a cutting-edge computational tool: self-supervised speech models, neural networks of the same family as HuBERT that learn rich representations of speech sound by being trained on massive amounts of audio without labels. Distances between talkers in the representational space of such models have been shown in prior work by the same group to capture perceptually meaningful acoustic similarity, providing a principled way to measure how alike any two speakers sound to a machine that has learned the statistics of human speech.

From this large dataset, the researchers constructed twenty training-test talker pairs, deliberately engineered so that the pairs spanned a wide range of intelligibility levels and acoustic similarity values. Each pair was then evaluated by ten native English listeners in a sentence-in-noise transcription task. The design was an exposure-test format: listeners first heard sentences produced by the training talker, giving their perceptual system a chance to adapt, and then transcribed sentences from the test talker, a different accented speaker they had not encountered before. Accuracy on those test sentences, measured as the proportion of words correctly transcribed against the noise, served as the index of cross-talker generalization. The inclusion of background noise was not incidental. Noise is one of the most ecologically common and cognitively demanding listening conditions, and prior work has shown that understanding accented speech in noise is substantially harder than in quiet, making any adaptation benefit both more meaningful and more measurable.

The headline result was genuinely encouraging: most of the twenty talker pairs showed cross-talker generalization. In other words, brief exposure to a single second-language talker was frequently sufficient to improve listeners’ comprehension of a completely different second-language talker. This finding matters practically. It suggests that single-talker training paradigms, which are far cheaper and simpler to implement than multi-talker regimes, can still deliver real-world listening benefits, whether in language education, in preparing listeners for communication in multilingual workplaces, or in clinical contexts where listeners must adjust to unfamiliar voices.

But then comes the twist that gives the study its title. When the researchers asked whether the size of the generalization effect was predicted by the intelligibility of the training talker, the answer was no. When they asked whether it was predicted by the acoustic similarity between training and test talkers, again the answer was no. Neither variable reliably drove the magnitude of transfer across the twenty pairs. The effects, in statistical terms, refused to line up with either theoretical account. The researchers had planned a replication strategy, intending to test additional talker triplets after the initial set, but because no clear patterns emerged in the first batch, the planned follow-up analyses were not pursued, and the untested triplets’ data were deposited in the project’s Open Science Framework repository for other researchers to examine.

This null result is more provocative than it might first appear. In science, a clean failure of two leading hypotheses is a map of where not to dig, and it forces the field to consider what else could be carrying the load. One possibility is that generalization depends on properties of the listener rather than the talker: individual differences in working memory, attention, or prior experience with accented speech are known to shape how people cope with difficult listening conditions, and such listener-side factors could swamp talker-side predictors in samples of ten listeners per pair. Another possibility is that the relevant similarity is not global acoustic distance but something more structured and linguistic, such as overlap in specific error patterns, for example whether two talkers share the same substitutions of particular English vowels or consonants. Two speakers could be acoustically distant overall yet confusable in exactly the phonemic categories that matter for transcription, a hypothesis that coarse representational distances might miss.

The study also carries a methodological lesson that extends well beyond speech perception. The use of self-supervised model representations to quantify acoustic similarity represents a broader shift in the cognitive sciences, where large-scale neural models trained on natural data are increasingly used as measurement instruments for psychological constructs. The same approach has recently been used by other groups, including work presented by Jin, Zhu, and Jaeger, to predict generalization of adaptation across talkers from latent speech representations. Yet the current findings temper the enthusiasm: even a sophisticated, data-driven similarity metric grounded in state-of-the-art speech models did not predict human generalization in this paradigm. The representations may capture acoustic similarity faithfully, but what the human perceptual system extracts from a brief encounter with one accented voice may be organized along dimensions that neither overall intelligibility nor global acoustic distance adequately describes.

What remains is a robust empirical anchor and an open question. The empirical anchor is that single-talker training works more often than not: listeners who adapt to one second-language voice typically carry some of that benefit to new voices, even under the adverse condition of background noise. The open question is the mechanism. The authors are candid that the predictors remain elusive, and that candor is itself a contribution, redirecting the field away from tidy two-factor stories and toward richer models that incorporate listener variability, linguistic structure, training design, and the temporal dynamics of perceptual learning. For the millions of people who navigate multilingual environments daily, the practical takeaway stands: a little exposure goes a long way. For the scientists, the accent on the horizon is clear even if its coordinates are not. The data and materials from the study are publicly available, ensuring that the next generation of hypotheses about how the brain learns to understand unfamiliar speech can be tested against the same evidence.

Subject of Research: Cross-talker generalization of perceptual adaptation to second-language speech after single-talker training

Article Title: Generalization of perceptual adaptation to second-language speech in a single-talker training paradigm: Predictors remain elusive

Article References: Kim, S.-E., Goldrick, M., Keshet, J., & Bradlow, A. R. (2026). Generalization of perceptual adaptation to second-language speech in a single-talker training paradigm: Predictors remain elusive. Attention, Perception, & Psychophysics, 88(7), Article 192. https://doi.org/10.3758/s13414-026-03333-5

Image Credits: AI Generated

DOI: 10.3758/s13414-026-03333-5

Keywords: speech perception, perceptual learning, second-language speech, cross-talker generalization, single-talker training, speech intelligibility, acoustic similarity, self-supervised representations, accented speech, speech-in-noise, bilingualism, auditory adaptation

The motivation behind the toolkit is rooted in a long-standing tension within the field. Laboratory speech, produced in word-reading or sentence-repetition tasks, gives experimenters precise control over variables and supports repeatability, which is why it has anchored theoretical work in speech perception and production for generations. But an accumulating body of evidence shows that listeners and speakers dynamically shape one another in everyday conversation, adapting to accents, talkers, and social-linguistic associations in ways that rigid laboratory tasks cannot capture. The classic observer’s paradox, articulated by William Labov in 1972, compounds the problem: people who know they are being recorded often adjust their speech, consciously or not, making genuinely naturalistic samples scarce and difficult to obtain.

The logistical hurdles do not end there. Field recordings demand extensive preparation, community relationships, and careful attention to the researcher’s positionality. Reading tasks systematically exclude speakers — such as many heritage bilinguals — who are conversationally fluent but may lack literacy in the language being studied. Funding constraints make it increasingly difficult to build and share open-access speech corpora, a burden that falls hardest on graduate students, early-career scholars, and researchers working on minoritized language communities. The authors argue that these barriers do not merely slow research down; they distort the theoretical record, because dominant models of speech perception and production were built on idealized, read speech and often treat real-world linguistic diversity as noise rather than as the phenomenon itself.

TAPA’s answer is to stitch together a chain of proven open-source components into one Python pipeline that can be installed with a single command and run either locally or in Google Colab, where free GPU access removes the need for any local setup. The journey begins with a YouTube audio downloader built on yt-dlp, which pulls the audio stream from any standard video URL and converts it to an MP3 file; users can equally supply their own local audio. Transcription is handled by OpenAI’s Whisper, which produces word-level transcripts with precise start and end timestamps. Speaker diarization — deciding who spoke when — relies on two pretrained neural models: Silero VAD, which detects stretches of actual speech, and Resemblyzer, which maps each detected segment to a 256-dimensional voice embedding in which clips from the same speaker cluster together. Those embeddings are then grouped to assign every transcribed word to its speaker.

From there, the pipeline turns to fine-grained phonetic measurement. The Montreal Forced Aligner produces phone-level boundaries linking the audio to the transcript, using an American English acoustic model with a dictionary-based proportional-timing fallback when alignment is unavailable. Vowel formants — the resonant frequencies F1 and F2 that define vowel identity — are extracted with Praat-parselmouth, using the Burg algorithm over the central portion of each vowel after trimming the edges to suppress coarticulation, with the maximum formant ceiling adjusted adaptively to each speaker’s pitch. Stop consonant voice onset time, the interval between a stop’s burst release and the onset of voicing, is measured by Dr. VOT, a neural model that predicts burst and voicing boundaries directly from the waveform. Fricative spectral moments — center of gravity, standard deviation, skewness, and kurtosis — are computed over the middle 80 percent of each fricative interval. The output is a set of CSV or JSON files containing per-token measurements and per-speaker summaries.

To demonstrate what the pipeline can do, the team pointed it at a demanding piece of real-world audio: the 90-minute 2016 U.S. presidential debate between Hillary Clinton and Donald Trump, complete with audience noise, podium reverberation, overlapping talkers, and the performative speech style of political theater. From that single recording, TAPA extracted nearly 33,000 acoustic segments — 19,855 vowels, 6,547 stops, and 6,372 fricatives. The case study focused on the two main candidates, yielding more than 7,000 monophthong tokens each for vowel-space analysis, alongside thousands of stop and fricative tokens per speaker. Plotting the F1-by-F2 vowel spaces revealed that Clinton’s vowel categories occupied a larger acoustic area than Trump’s, and comparing both speakers’ category means against the classic Hillenbrand reference norms showed the expected centralization of vowels in continuous speech relative to citation-form productions.

Crucially, the team did not simply trust the automated output. A phonetically trained research assistant hand-coded stratified samples of vowels, stops, and fricatives from the debate audio, blind to the pipeline’s results, and the two sets of measurements were compared token by token. Vowel formants emerged as the pipeline’s strongest suit: agreement with expert measurements was high, with correlations of 0.89 for F1 and 0.87 for F2, mean absolute errors of 37 and 125 hertz respectively, and no meaningful bias for F1. Fricative measurements showed a more differentiated picture. Spectral standard deviation agreed strongly with hand coding overall, and center of gravity tracked the human reference closely for the sibilant fricatives — correlations of 0.87 for /s/ and 0.94 for /ʃ/ — while the spectrally indistinct non-sibilants /f/ and /θ/ agreed far more weakly, a pattern consistent with long-standing findings that these sounds are hard to classify from spectral moments alone. Clinton’s /s/ center of gravity sat roughly 900 hertz above Trump’s, echoing documented patterns of gender-linked variation in sibilant production.

Stop consonants told a more cautionary tale. Although the aggregate voice-onset-time means for voiceless stops /p/, /t/, and /k/ landed squarely in the expected American English long-lag range of roughly 69 to 83 milliseconds, the per-token picture was far less reassuring. The Dr. VOT model, which was trained on isolated single-word productions with stop classes artificially rebalanced to a 50:50 positive-to-negative ratio, mislabeled 45 to 61 percent of voiceless onset stops as prevoiced — an outcome incompatible with English phonology — and its per-token magnitudes correlated essentially not at all with expert measurements (r = −0.04), diverging by 50 to 80 milliseconds on average. The authors attribute this to a training–deployment mismatch: continuous, noisy debate audio with a natural 95:5 ratio of positive to negative stops violates both of the model’s core training assumptions. Rather than hide the problem, they report it transparently, omit voiced stops from the case study entirely, and recommend that aggregate means be treated as descriptive rather than per-token measurements.

That candor reflects the paper’s broader stance on automation in science. The authors are explicit that TAPA is meant to aid researchers, not replace them: acoustic training must continue, research must be deployed and interpreted by humans, and the energy costs of training and fine-tuning large models warrant critical scrutiny. The pipeline inherits the biases of its components — the default configuration is American English-centric, diarization degrades on heavily overlapping speech, and forced alignment stumbles on out-of-vocabulary names — so the team recommends that every substantive deployment include a small hand-coded validation subsample from the researcher’s own corpus, and that publications report which component produced each measurement. With those safeguards in place, TAPA offers something the field has lacked: a free, modular, inspectable path from a web video or field recording to speaker-level, phone-level acoustic data at a scale that would once have required years of manual annotation, potentially opening naturalistic speech research to communities and questions that laboratory methods have long left on the periphery.

Subject of Research: An open-source computational pipeline for automated acoustic-phonetic analysis of naturalistic multi-speaker speech recordings

Article Title: Toolkit for acoustic–phonetic analysis of naturalistic speech data

Article References: Kutlu, E., Peters, E., Tapanes, C., Chandio, S., & Khalid, O. (2026). Toolkit for acoustic–phonetic analysis of naturalistic speech data. Behavior Research Methods, 58(10), Article 289. https://doi.org/10.3758/s13428-026-03174-y

Image Credits: AI Generated

DOI: 10.3758/s13428-026-03174-y

Keywords: acoustic-phonetic analysis, naturalistic speech, speaker diarization, forced alignment, vowel formants, voice onset time, fricative spectral moments, open-source software, speech perception, sociophonetics, automatic speech recognition, Behavior Research Methods