RESEARCH DOSSIER
Auditory Scene Analysis
Abstract. Auditory scene analysis is the perceptual organization of acoustically superposed events into auditory objects and temporally coherent streams. This research monograph examines the scientific history of the problem; acoustic and perceptual distinctions between sources, components, streams, and objects; primitive and learned grouping cues; harmonicity, onset synchrony, spatial information, temporal coherence, repetition and continuity; sequential streaming dynamics; competing-speech selection; neural representations; predictive and computational models; development; musical organization; and conditions of failed segregation. The methodological approach is a critical narrative synthesis of original psychophysical experiments, human neural recordings, established theoretical works, and recent reviews, with attention to competing explanations and limits of inference. The analysis finds convergent support for multiple interacting grouping constraints and attention-dependent selection, while the number and location of mechanisms, causal hierarchy, ecological generalization, individual differences and artificial-system comparability remain open. No original experiments or meta-analysis are claimed.
Keywords: auditory scene analysis; auditory grouping; auditory stream segregation; perceptual organization; temporal coherence; harmonicity; common onset; auditory objects; auditory attention; cocktail party problem; informational masking; auditory cortex; speech perception in noise; auditory prediction; perceptual multistability; sound textures; musical scene analysis; computational auditory scene analysis; Artificial Era
Introduction: Hearing Organized Worlds
Auditory scene analysis (ASA) is the family of perceptual processes by which a listener organizes complex acoustic input into events, streams, and auditory objects. At a gathering, a listener may follow one voice, notice a glass breaking, and experience music as background even though the pressure wave at each ear combines contributions from all these sources. The scientific problem arises because addition of acoustic waves is physically straightforward while attributing their components to causes is an inverse inference problem. An individual acoustic component is not already labeled by its physical producer. Perceptual organization must establish which components belong together at a moment and which events persist together through time Bregman (1990) Carlyon (2004).
This monograph asks how acoustic regularities support the segregation of concurrent sounds, how sequential events become streams, how attention interacts with grouping, how neural activity expresses alternative organizations, and why selection fails under some listening conditions. It also asks which conclusions survive transfer from pure-tone laboratory paradigms to speech, music and natural soundscapes. The evidence base comprises original behavioral experiments, electrophysiological and human cortical studies, books that established the framework, computational models, and critical reviews including the recent study of musical scene analysis Szabó et al. (2016) Siedenburg et al. (2026).
The argument proceeds from a measurement distinction: acoustical decomposition, subjective reports of streams, recognition of source identity, and comprehension of a spoken message are different dependent variables. A person can identify a voice without correctly reporting every competing component, or hear two streams without determining two physical sources. Likewise, a computational separator can recover waveform estimates without reproducing human organization. The goal is a disciplined account of perceptual grouping that preserves these explanatory levels and assesses the conditions under which experiments connect them Winkler et al. (2009) Shamma et al. (2011).
Scientific Definitions: Source, Event, Stream, and Object
A physical sound source is an event-producing system in the environment; an acoustic signal is a pressure fluctuation measured at a location; an auditory event is a perceptually bounded occurrence; and an auditory stream is a perceptual grouping of events through time. The term auditory object can refer to a stable source representation or to an organized pattern such as a melody, depending on the theoretical tradition. The mapping is many-to-many: a violin's note components can split perceptually; several instruments can blend into one expressive musical texture. Such cases are fundamental tests of any definition that equates perceived stream count with physical source count Bregman (1990) Szabó et al. (2016) Siedenburg et al. (2026).
Simultaneous organization groups frequency components arriving together, commonly described as deciding which partials form the current auditory entity. Sequential organization links events separated in time, potentially creating distinct temporal streams. The processes interact: an asynchronously starting harmonic may acquire a distinct identity, and a preceding tone may subsequently pull one component away from a concurrent complex. Laboratory reports commonly use 'one' versus 'two' streams, yet a richer representation must also specify temporal predictability, partial fusion, attentional accessibility, source localization, and the possibility of multiple defensible interpretations Darwin & Ciocca (1992) Tougas & Bregman (1990).
The expression 'auditory scene' is thus an explanatory construction about coherent environmental causes and patterns. It is not the raw spectrogram, a complete acoustic reconstruction, or a fixed conscious picture. Researchers observe choices, reaction times, detection thresholds, discrimination, identification, and neural responses; latent objects and streams are inferred through controlled comparisons. Psychophysical construct validity depends on showing that task instructions and criterion shifts do not wholly explain a reported organizational change. This matters most for ambiguous stimuli whose physical sequence remains constant while perception switches Winkler et al. (2012) Snyder & Alain (2007).
The Historical Emergence of the Scene-Analysis Problem
Experimental attention research gave the modern scene problem an early operational form. In 1953 E. Colin Cherry presented simultaneous speech and investigated recognition when messages were delivered to one or both ears. The question was how a listener isolates and understands an intended message in a competing acoustic field. Dichotic listening thereby became a way to manipulate selection, memory, and informational competition, though its conditions do not capture all the spatial and reverberant properties of everyday multispeaker environments Cherry (1953).
Research on sequential tones supplied a complementary tradition. Alternating patterns can be heard as an integrated sequence or as two frequency-specific streams; studies of temporal spacing, pitch separation and adaptation made organization measurable independently of lexical comprehension. Anstis and Saida examined adaptation effects in repetitive sequences, and the field developed the 'build-up' concept to describe increasing likelihood of segregation during repeated listening. Later evidence made clear that the temporal course of first reports, ongoing perceptual states, and prior stimulation must be analyzed separately rather than treated as a single accumulating mechanism Anstis & Saida (1985) Winkler et al. (2012).
Albert S. Bregman's 1990 monograph unified the research problem across speech, music, psychophysics, and perceptual organization. He distinguished relatively stimulus-driven grouping from schematic, experience-informed processes and treated auditory analysis as the recovery of structured descriptions from mixtures. This synthesis gave the field its enduring name and an experimentally fertile vocabulary. Subsequent work extended that account to representations of regularities, cortical responses, attentional modulation, computational inference and multistability. Historical credit belongs to several converging traditions rather than to one laboratory or one putative grouping law Bregman (1990) Winkler et al. (2009) Micheyl et al. (2007).
The Inverse Problem and the Logic of Grouping Cues
At a microphone or eardrum, pressure contributions sum. From the resulting waveform, an observer could in principle posit innumerable decompositions unless constrained by structure in the source-generating world. Auditory grouping is informed by such constraints as harmonic relations among partials, common temporal onset, shared frequency or amplitude changes, coherent trajectories, and consistent spatial attributes. These features provide evidence of common origin because sound-producing bodies often vibrate with correlated spectra and temporal envelopes. Their diagnostic value is probabilistic and context-sensitive rather than an infallible source label Darwin (1997) Shamma et al. (2011).
The cues are also partly redundant and partly in conflict. Harmonicity can promote fusion while a leading or asynchronous component promotes separation; common modulation may mitigate certain harmonic mistuning effects. A theory that specifies only a list of good grouping cues cannot predict how these competing constraints are weighted when they disagree. Controlled conflicts therefore have greater explanatory leverage than illustrations where all cues point toward the same conclusion. The experiments of Darwin and Ciocca, for example, operationalize these interactions through mistuned harmonic components, onset timing and ear of presentation Darwin & Ciocca (1992).
Spatial location can help maintain a selected voice, but the perceptual advantage of interaural differences depends on the task, frequency resolution, and overlap of competing signals. Darwin's review emphasized that the cues useful for localizing a source need not suffice to disentangle simultaneous partials of that source from similar competitors. Auditory research accordingly distinguishes the physical separability of waveforms, segregability of perceptual components, spatially supported selection, and assignment of source identity. These distinct functions often co-occur, but their dissociations reveal the organization process Darwin (1997) Kidd et al. (2005).
Research Logic, Methods, and Evidentiary Hierarchies
Auditory scene analysis is studied through carefully controlled sound synthesis, psychophysical reports, behavioral performance and neural measurements. A minimal paradigm manipulates one acoustic parameter, such as pitch distance between alternating tones, while holding presentation level, rhythm, and trial structure as constant as possible. A richer paradigm presents concurrent speech or natural environmental sounds and measures recognition or comprehension. Investigators should distinguish a change in whether sounds are perceptually separated from a change in a listener's willingness or ability to report separation Carlyon (2004) Szabó et al. (2016).
Three kinds of claim require different warrants. A discrimination experiment establishes sensitivity under particular listening conditions; a multistable streaming report demonstrates that identical stimuli can admit multiple perceptual organizations; and cortical recording establishes an association between neural dynamics and task or perceptual state. Strong causal mechanistic claims need interventions, convergent behavioral controls, and models that predict new outcomes. The auditory cortex is implicated by many findings, yet the stage at which a perceptual organization becomes effective cannot be read directly from the earliest neural correlate of acoustic contrast Micheyl et al. (2007) Cusack (2005).
This is a critical narrative research synthesis, not a preregistered exhaustive systematic review or statistical meta-analysis. Source selection privileges landmark original experiments, verified publication records, direct-access full texts when available, and recent reviews capable of exposing disputed conclusions. No pooled effect size, prevalence figure, or numerical mechanism parameter is inferred from heterogeneous papers. When an original study is available only as a bibliographic record or abstract, conclusions remain at the reported level. Evidence quality is evaluated through controls, ecological scope, representation of participants, stimulus construction, replicability, and alternative accounts Calcus (2024) Snyder & Alain (2007).
Sequential Streaming: Frequency Separation and Rate
Alternating-tone experiments isolate sequential grouping by cycling sounds of frequencies A and B in patterns such as ABA followed by a pause. At small separations or slow presentation rates, listeners commonly perceive a single, temporally ordered pattern. As separation or speed increases, two streams become more likely, one carrying A events and another B events. Crucially, segregation alters the availability of cross-frequency temporal order: in strongly separated streams the listener may track two independent rhythms more readily than the exact interleaving of A and B. The underlying acoustic sequence has not changed; its perceptual assignment has Bregman (1990) Carlyon (2004).
Frequency difference is often expressed as an interval relative to a center frequency. Such normalizations assist comparison, but pitch spacing, peripheral excitation and frequency-channel selectivity are not interchangeable descriptions. Stream reports depend on duration, sound level, regularity, instructions, response format and the prior exposure of a listener. A psychophysical experiment needs to distinguish one-stream predominance from the absence of capacity to resolve frequencies, because a listener can discriminate isolated tones yet group them in a common sequence. Neural and perceptual interpretations must also account for frequency-specific stimulus effects that arise before perceptual organization Snyder & Alain (2007) Micheyl et al. (2007).
Although this paradigm is minimal, its advantages are substantial: rapid timing and frequency parameters can be varied systematically, ambiguous states can be sampled repeatedly, and neural responses can be compared to direct perceptual reports. Its principal limitation is the absence of many features of natural sound sources, including complex harmonic structure, reverberation, semantic meaning and competing goals. Therefore, success in predicting the perceptual boundary of ABA sequences is one component of a theory of general auditory scenes rather than evidence that natural speech segregation has been solved Szabó et al. (2016) Shamma et al. (2011).
Build-Up, Adaptation, and the Timing of Perceptual Decisions
The term build-up refers to the common observation that two-stream reports become more frequent over the course of repeating ambiguous sequences. Anstis and Saida's adaptation experiments were historically influential because they showed how preceding acoustic history modifies later grouping. However, the usual aggregate curve can combine several processes: listeners require time to reach an initial report, their report criterion can shift as a sequence continues, and their currently perceived organization can alternate. A trial-averaged increase in segregation is not itself the time constant of a uniquely identified sensory adaptation mechanism Anstis & Saida (1985) Winkler et al. (2012).
Physiological explanations often invoke differential activity in frequency-tuned neural populations, adaptation or stimulus-specific suppression, and competition between alternative encodings. Attention-based explanations consider whether time and task engagement allow the relevant grouping hypothesis to stabilize. The neglect and divided-attention studies by Carlyon and colleagues found substantial reductions in build-up when listeners directed attention elsewhere, showing that an adequate account must include attentional conditions. Yet the presence of attentional modulation does not imply that every elementary grouping computation requires focused conscious monitoring Carlyon et al. (2001) Micheyl et al. (2007).
The methodological lesson is to analyze individual onset latencies, percept durations, transitions and response delays separately. Reports made at a few fixed times can create distorted build-up functions; continuous tracking introduces its own motor and criterion effects. Repetition across trials can condition the observer and induce adaptation or expectations. Stimulus-specific and perception-specific aftereffects need controls that dissociate what was heard from what was presented, and sequential priming should not be conflated with evidence for a universal default integrated percept Winkler et al. (2012) Denham & Winkler (2006).
Simultaneous Grouping: Harmonicity and Onset Synchrony
Complex periodic sounds contain harmonically related components that often fuse into a single pitch-bearing auditory event. When one component is shifted away from the harmonic series, listeners may hear it as a separate tone or find its contribution to the compound pitch diminished. The classic grouping experiments by Darwin and Ciocca used mistuning, onset asynchrony and dichotic presentation to investigate these effects. Such manipulations matter because they selectively challenge the assignment of a component to a common source while preserving much of the stimulus Darwin & Ciocca (1992) Darwin (1997).
Common onset is another strong grouping cue: spectral components that begin together often support a single event, while a component that begins earlier may continue as its own stream when overlapping components join later. Harmonicity and onset cues may conflict. The expected consequence depends on relative cue strength, stimulus duration and the perceived continuity of the earlier event. Grouping effects cannot be equated simply with masking, because a component may remain physically present and individually detectable while its assignment to the heard complex changes. This is why studies pair pitch judgments with detectability and component-selection controls Darwin & Ciocca (1992) Tougas & Bregman (1990).
A useful generalization is that grouped components tend to share lawful change, not merely an initial temporal coincidence. Components that vary in a coordinated manner may be inferred to belong to a single changing source. Darwin's work on common modulation showed that this inference depends on modulation type and the relevant spectral structure; shared frequency modulation and shared amplitude modulation need not exert the same effect in pitch grouping. The cue vocabulary must thus be expressed as testable hypotheses about distributions of acoustic relationships rather than as a list of universal laws Darwin (1997) Shamma et al. (2011).
Temporal Coherence, Common Fate, and Cross-Frequency Binding
Temporal-coherence models propose that distributed auditory features can be bound when their activity changes in coordinated fashion over time. Rather than requiring source-specific labels on every feature channel, the model asks whether envelopes and patterns exhibit correlated fluctuations that plausibly share an environmental cause. This hypothesis offers a possible common principle connecting simultaneous organization with sequential continuity and permits attention to alter which coherent relations are accentuated. It is articulated explicitly in the review by Shamma, Elhilali and Micheyl Shamma et al. (2011).
Elhilali and colleagues experimentally investigated temporal coherence in relation to both perceptual organization and cortical representation, using controlled sound patterns that probed how coherent changes promote grouping. The experiments supported the relevance of temporal relations beyond static separation of frequency channels. Their importance lies in a manipulable cue structure rather than in proving that one mathematical coherence statistic explains every auditory source. Different time scales, peripheral filtering, acoustic modulation and attentional state can change which correlated fluctuations are actually available to the perceptual system Elhilali et al. (2009).
Several explanatory commitments remain unsettled. A temporal-coherence measure can be predictive while leaving open where its computation occurs, whether attention sets a template before grouping or selects among completed groups, and how learned expectations reshape the relevant features. A coherent environmental sound may display internal asynchronies; two independent sources can occasionally share a rhythm. Strong evaluation therefore uses purposely conflicting cues, noise and reverberation controls, source reconstruction tasks, and measures that distinguish a listener's perceptual assignment from simple recognition of rhythmic similarity Shamma et al. (2011) Szabó et al. (2016).
Auditory Continuity, Occlusion, and Restoration
In natural environments sound sources are interrupted by obstacles and competing louder events. An auditory stream may nonetheless be perceived as continuing through a masking interval, provided the surrounding acoustic evidence supports that interpretation. Continuity and restoration effects show that the listener forms an organized account of a possible ongoing source rather than registering only the momentary audibility of each component. The perceived event can outlast the presence of recoverable source energy in the mixture Bregman (1990) Tougas & Bregman (1990).
Tougas and Bregman's experiment directly compared auditory stream organization under conditions involving a continuity illusion. Their reported result was that streaming organization tended to be similar whether or not perceptual restoration occurred, alongside further effects of frequency proximity and harmonic enrichment. This finding constrains models that treat perceptual completion as a necessary intermediate stage for all stream grouping. The experiment establishes an informative dissociation for its stimuli; it does not claim that continuity and streaming never interact under more complex or longer-lived acoustic scenes Tougas & Bregman (1990).
Restoration should be distinguished from veridical reconstruction. A listener's impression that a tone or speech segment continued under a louder sound does not supply the missing waveform. The inference is compatible with multiple possible masked events. Experiments can vary masker plausibility, continuity of pitch trajectory, onset and offset transitions and the acoustic context, then compare detection, recognition and phenomenological reports. Such designs make the distinction between acoustic evidence and perceived completion central to the science of auditory objects Bregman (1990) Winkler et al. (2009).
Multistability and the Limits of a Single Acoustic Partition
Ambiguous auditory sequences can support alternating percepts without any change in stimulus. In classic streaming tasks the same repeated tones may appear momentarily as one pattern, then as two separate sequences, and occasionally return to integration. Such multistability shows that the inference from acoustic events to perceptual objects is not a deterministic labeling of components. It provides a special test for models because physical input can be fixed while perceptual state is observed over time Winkler et al. (2012) Denham & Winkler (2006).
Predictive and competitive accounts posit alternative organizations that remain available and interact through adaptation, prediction, stochastic fluctuations or attentional selection. Winkler and colleagues interpreted multistability within a predictive-coding framework: regularities generate expectations about future events, and alternative stream interpretations compete for expression in awareness. This is a useful explanatory model, but evidence of alternation alone cannot establish exactly which neural populations carry predictions or whether one hierarchical architecture is uniquely correct Winkler et al. (2012) Winkler et al. (2009).
Research designs must account for report dependence. A button press for 'one stream' or 'two streams' is delayed relative to the internal transition and may be influenced by observer instructions. No-report physiological indicators, repeated stimulus segments, and independent tasks can reduce some ambiguities, yet those approaches introduce their own proxy assumptions. Robust work reports distributions of percept durations and transitions, checks whether percepts correlate with acoustic history, and tests predictions that distinguish causal models rather than merely reproducing a familiar switching curve Snyder & Alain (2007) Szabó et al. (2016).
Selective Attention: Tracking One Organization in a Mixture
Selective listening is a capacity to prioritize information associated with a relevant sound source while other acoustic events continue to reach the ears. Cherry's speech experiments established the enduring task of recognizing one message amidst competing speech, and later studies asked how unattended messages affect behavior and memory. Selection can occur after some acoustic organization has segregated candidate objects, but expectations about a target voice can also shape the organization achieved. Attentional selection and grouping therefore interact over time and cannot be assigned a single universal order from behavioral performance alone Cherry (1953) Wood & Cowan (1995).
A basic distinction concerns selecting a location, selecting an auditory feature, and selecting a perceptual object. A spatial target can carry several concurrent sounds; a pitch feature can recur across different physical sources; a voice can move while remaining perceptually identifiable. If target cues change, listeners may maintain continuity through learned timbre, speech rhythm, location and higher-level meaning. Experimental cueing designs should independently vary these attributes rather than use a single predictable talker and infer an object-based mechanism from accurate report Darwin (1997) Carlyon (2004).
Attention can modify perception without explaining all preattentive regularity processing. Carlyon and colleagues measured auditory stream segregation under attention manipulations, including unilateral neglect, and found evidence that attention contributes to the temporal development of segregation. Such results challenge a simple two-stage account in which auditory streams are completed independently before attention begins operating. Yet the degree and timing of attentional necessity differ among stimuli, measures and ongoing task demands; the relevant inference is an interaction, not a universal exclusive requirement Carlyon et al. (2001) Shamma et al. (2011).
The Cocktail Party Problem and the Distinction Between Masking Mechanisms
The cocktail party problem encompasses at least two kinds of competition. Energetic masking occurs when overlapping acoustic energy limits access to target cues in the auditory periphery or subsequent frequency analysis. Informational masking concerns interference associated with similarity, confusion, uncertainty about target identity or failure to select the intended stream despite physically available cues. In speech experiments the categories are operational aids; their quantitative partition depends on masker construction and listener performance measures Kidd et al. (2005) Cherry (1953).
Kidd, Mason and Gallun studied how energetic and informational masking combine during speech identification, providing evidence that interference cannot be predicted from acoustic overlap alone in the tested conditions. The main interpretive advantage of experiments separating the two is that they test different explanatory bottlenecks. A spatially separated interferer may become less confusable without an equivalent reduction in its physical energy, while a filtered masker may reduce energetic overlap without eliminating the difficulty of identifying a target among competing linguistic patterns Kidd et al. (2005).
Speech identification depends on temporal envelope information, phonetic structure, linguistic knowledge and attention. A listener's apparent ability to segregate voices may improve with familiarity or predictable content even while acoustic separation remains unchanged. For this reason, claims about 'scene analysis' inferred from word-recognition accuracy need accompanying tests of source assignment or perceptual continuity. The target may be intelligible because context supplies missing words, because the masker becomes less potent, or because perceptual grouping improves; those possibilities require different controls Wood & Cowan (1995) Zion Golumbic et al. (2013).
Object Selection, Memory, and Unattended-Speech Claims
Selective-attention paradigms often ask listeners to shadow a designated message and later report information delivered on the unattended channel. Cherry's observations have sometimes been retold as proof that unattended speech is never analyzed. The experiment and subsequent revisitations support a more qualified account: performance and later memory depend on the task, perceptual salience and the kind of unattended feature tested. Behavioral failures to recall a message are not direct measurements of whether its acoustic, phonetic or semantic information was neurally represented Cherry (1953) Wood & Cowan (1995).
Wood and Cowan revisited the classic procedure and demonstrated that attention and memory claims change when critical unattended events and reporting procedures are examined carefully. Their study underscores that awareness, retrospective report, immediate distraction and persistent memory traces are separable dependent variables. An unexpected event may redirect attention, while a participant may forget the event's details despite having detected it at the time. The theoretical object under investigation must be named precisely: entry into awareness, attentional capture, continuing stream selection, or later recollection Wood & Cowan (1995).
This distinction bears directly on experimental inference. Investigators should compare tasks using online detection, probe recognition, stream-tracking reports and delayed recall; they should also report response criterion and trial probabilities. A single surprising unattended word or name has different processing requirements from continuous comprehension of an ignored sentence. The term 'unattended' describes the experimental instruction and the selected focus, not proof of zero processing outside that focus Wood & Cowan (1995) Shamma et al. (2011).
Speech Tracking in Human Auditory Cortex
Human neural recordings allow research to ask how the brain follows a chosen speech stream when multiple voices are physically present. Mesgarani and Chang recorded responses from human superior temporal cortical sites during multitalker perception and demonstrated selective cortical representation of the attended speaker. This was a major advance because it related neural responses to the competing linguistic signals themselves rather than treating an overall increase in cortical activity as evidence of attention Mesgarani & Chang (2012).
Zion Golumbic and colleagues examined mechanisms underlying selective neuronal tracking of attended speech at a cocktail party. Their work connected ongoing neural signals to the temporal structure of attended versus unattended speech, sharpening the hypothesis that neural activity follows behaviorally relevant acoustic features in a competition. Neural tracking, however, admits multiple mathematical explanations: common stimulus rhythm, top-down gain, task-dependent prediction and changes in acoustic signal-to-noise ratio can produce different components of the measured response Zion Golumbic et al. (2013).
Encoding and decoding models must be validated on held-out speech materials, participants and attentional conditions before they support strong claims about real-world source separation. The finding that a classifier can distinguish which speaker is attended does not itself show that cortical activity reconstructs both clean original waveforms. Speaker selection, intelligibility, perceptual organization and decoding accuracy must be evaluated through independent behavioral outcomes. This distinction is equally important when comparing biological recordings to artificial neural speech separators Mesgarani & Chang (2012) Zion Golumbic et al. (2013).
Cortical Populations, Adaptation, and Stream Formation
Auditory cortex exhibits responses shaped by stimulus frequency, repetition, temporal modulation and attention. Micheyl and collaborators reviewed evidence relating auditory cortical responses to perceptual streaming, including the possibility that adaptation within frequency-selective populations contributes to the separation of alternating tone sequences. Such physiological findings are important because they connect controlled stimulus manipulations to candidate neural mechanisms. Yet a neuron distinguishing the A and B tones can do so without encoding a perceptually segregated scene; frequency selectivity is a necessary substrate for many tasks, not a sufficient explanation of perceptual organization Micheyl et al. (2007).
Snyder and Alain surveyed electrophysiological and behavioral studies toward a neurophysiological theory of stream segregation. The synthesis emphasized both cortical sensitivity to acoustic contrast and the temporal evolution of organization. For claims that a neural marker indexes an actual stream percept, experiments must compare acoustically matched conditions in which reports differ, or show that an intervention shifts the perceptual organization while relevant physical cues are held stable. Even then, report-related, adaptation-related and decisional neural signals can overlap Snyder & Alain (2007) Winkler et al. (2012).
A complete account likely spans subcortical feature coding, cortical regularity representations, higher-order selection and frontal-parietal contributions. Cusack associated perceptual organization with the intraparietal sulcus in a human imaging study, providing an example of how the scene-analysis problem extends beyond primary auditory representations. Localization of activation identifies a participating network under particular task conditions rather than an exclusive 'stream center.' Causal localization would require convergent timing, perturbation and model-based evidence Cusack (2005) Micheyl et al. (2007).
Spatial Scene Structure and the Limits of Laboratory Generalization
The sound reaching two ears often differs in level, phase and time of arrival. Such interaural differences assist localization and can help a listener distinguish sources, but the information becomes less stable when room reflections, movement and multiple talkers interact. Spatial separation can reduce some kinds of informational competition and improve speech recognition under certain configurations. Its role should be analyzed with spectral, harmonic and temporal cues because location alone does not define the grouping process Darwin (1997) Kidd et al. (2005).
Laboratory stimuli often use headphones, steady sounds and artificial frequency patterns. This control enables inference about particular grouping cues but suppresses natural complexity: source motion, reverberant tails, changing microphone geometry, overlapping music and variable background noise. A scene-analysis mechanism that correctly predicts frequency-based streaming may nevertheless fail when tested on transient environmental sound or conversational turn-taking. The appropriate progression is from identifiable acoustic causes to richer mixtures, with measured noise conditions, reproducible sound files and listener-relevant tasks Carlyon (2004) Szabó et al. (2016).
Ecological validation should preserve explicit performance metrics and the difference between recognition and source reconstruction. One design can combine a source-count judgment, source-location estimate and intelligibility test for the same mixture. Another can compare repeated listening under changing room acoustics with acoustically matched synthetic controls. The strongest generalization comes from models predicting all these outcomes under held-out mixtures, rather than from matching a handful of classic textbook illusions Shamma et al. (2011) Snyder & Alain (2007).
Prediction, Regularity, and Learned Auditory Expectations
Auditory organization depends partly on lawful relationships extending beyond a single acoustic frame. An alternating sequence creates expectations about the pitch and timing of upcoming tones, and predictable recurrence allows one candidate organization to anticipate events more successfully than another. Winkler, Denham and Nelken argued that regularity representations and predictions are central to auditory objects. The theoretical value of this proposal is its attempt to explain both grouping and changes in organization as a function of learned acoustic structure Winkler et al. (2009).
Denham and Winkler developed an account linking predictive representations to the formation of auditory streams. The model addresses why sound sequences can settle temporarily into distinct coherent descriptions, why their interpretations change and how prior auditory context matters. The hypothesis creates testable contrasts between the processing of unexpected events in an established stream and the processing of physically identical events when the regularity is different. However, an observed mismatch response or prediction effect does not automatically isolate the locus of perceptual grouping; stimulus adaptation and attention can contribute Denham & Winkler (2006) Winkler et al. (2012).
A scientifically useful predictive account must specify what is predicted, which measurements reveal the prediction, and how rival hypotheses are updated. Predictions may concern onset time, frequency, timbre, source identity or sentence-level linguistic structure. Models linking all these predictions under a common label require explicit interfaces across time scales and levels of representation. Otherwise 'prediction' risks serving as an unconstrained explanation for any change in performance. The evidence favors expectation-sensitive organization while leaving open the architecture and degree of generality of its computations Winkler et al. (2009) Szabó et al. (2016).
Computational Models of Segregation and Auditory Objects
Computational models of scene analysis range from signal-processing algorithms based on harmonic structure and modulation to neural competition, temporal coherence, probabilistic inference and predictive representations. Szabó, Denham and Winkler reviewed these model families and emphasized that simulations can clarify which assumptions are sufficient to reproduce particular streaming observations. A model of perceptual grouping should take acoustic stimuli as input and predict specific behavioral observables, including response distributions, switching, integration and errors, rather than merely drawing attractive spectrogram partitions Szabó et al. (2016).
One modeling strategy first extracts candidate acoustic features from the waveform and then groups features by local similarity and temporal coherence. Another uses multiple generative hypotheses about sound-producing events and estimates which hypothesis best explains the mixture. Both require choices about representation, priors, regularization and time scale. Differences in model performance can arise from feature extraction and optimization rather than from their claimed perceptual principles. Accordingly, comparable inputs, matched complexity and predeclared evaluation measures are necessary for a strong scientific comparison Shamma et al. (2011) Szabó et al. (2016).
Reproducing a classic build-up or frequency-separation curve supplies a constrained success, but should be complemented by stress tests. A candidate model should handle simultaneous grouping conflicts, stream switching under identical stimuli, temporal continuity under masking, attention changes, and transfer to unseen sound classes. Validating on original experiments and independent replication stimuli reduces overfitting to a benchmark. Neural plausibility is a separate criterion from behavioral fit; a model can predict correct choices while implementing a mechanism that is inconsistent with cortical recording or perturbation evidence Micheyl et al. (2007) Winkler et al. (2012).
Natural Sounds, Repetition, and the Discovery of Embedded Sources
Everyday soundscapes contain repetitive but partly hidden structure: footsteps within traffic, recurring calls in a forest or repeating machinery under ventilation noise. McDermott, Wrobleski and Oxenham studied recovery of sound sources from embedded repetition, showing that repetition can reveal acoustic components initially obscured by a mixture. This type of task is valuable because it introduces structured learning and sound recognition beyond alternating-tone stimuli. Its outcome depends on repeated exposure, mixture composition and the structure of the target rather than on a general ability to invert arbitrary acoustic addition McDermott et al. (2011).
Repetition can support more than one psychological operation. It may improve sensitivity to a target signal, establish its temporal template, enhance detection of deviations or permit a listener to recognize a source only once its characteristic structure has been learned. Distinguishing these operations requires comparing novel repeated sources to familiar ones, manipulating whether the target or background repeats, and testing whether recovered content transfers to a new mixture. Without these controls, improvement with repetition need not be evidence of a distinct auditory object representation McDermott et al. (2011) Winkler et al. (2009).
Naturalistic tasks also raise questions about perceptual completion and source identity. A sound may be identified as an engine or a bird while a complete isolated waveform remains unavailable. Reconstructive algorithms, source-label decisions and listener reports should therefore be treated as related but independent endpoints. Psychophysical experiments that measure all three can better discriminate between models of auditory inference and models of pattern recognition. The importance of natural sound repetition is that it provides an intermediate complexity level between simple tones and uncontrolled everyday scenes Bregman (1990) McDermott et al. (2011).
Sound Texture and Statistical Representation of Crowded Environments
Not every auditory environment is experienced as a list of distinct sources. Rain, applause, a stream or a densely mixed crowd may be represented as a texture characterized by statistical regularities. McDermott and Simoncelli synthesized textures by matching statistical properties of model auditory representations, demonstrating that certain natural-sounding textures can be generated from carefully selected peripheral-inspired statistics. This provides a scientific counterpoint to models that assume the goal of hearing is always to recover each source separately McDermott & Simoncelli (2011).
McDermott, Schemitsch and Simoncelli later tested how listeners discriminate sound textures over different durations. They reported opposing duration effects depending on whether listeners compared different textures or different exemplars of the same texture, consistent with a transition from access to local details toward stable summary statistics. The experiment indicates that time-averaged statistics are behaviorally relevant for some sound categories. It does not mean that all environmental scenes are represented only by texture summaries, or that every individual source is discarded once a texture forms McDermott et al. (2013).
These findings motivate a richer taxonomy of scene organization. A listening system can track selected foreground objects while summarizing many unattended background events statistically; it can also shift between these modes as the task changes. The boundary between object and texture should be investigated using controlled manipulations of event density, temporal persistence, structural repetition and reporting requirements. A theory that accommodates both identifiable sources and diffuse acoustic ensembles better matches the range of human auditory experience than a universal clean-separation ideal McDermott & Simoncelli (2011) McDermott et al. (2013).
Development of Auditory Organization Across Childhood
Auditory grouping and attentional selection develop through interacting sensory, cognitive and neural changes. Calcus reviewed behavioral and physiological evidence on how children analyze mixtures and attend to targets amidst distractors. Early capacities for segregation are present in infancy, while the efficiency and flexibility of complex listening continue developing into adolescence. The developmental trajectory is therefore not adequately summarized as a single age at which 'scene analysis appears'; different cues and tasks mature on different courses Calcus (2024).
Age comparisons require careful control for peripheral hearing sensitivity, auditory filter development, memory span, linguistic familiarity, and understanding of instructions. A child's lower speech-in-noise score may reflect limited lexical knowledge or response demands even when an elementary harmonic grouping process functions well. Conversely, successful detection of simple alternations may coexist with substantial difficulty maintaining a target speech stream. Developmental designs should include nonverbal measures where possible and distinguish immediate segregation from sustained attention Calcus (2024) Carlyon et al. (2001).
Longitudinal and intervention evidence could clarify which processes improve with experience. Tasks could compare changes in harmonicity use, temporal-coherence sensitivity, streaming build-up and attentional selection in the same observers across development. Statistical models would need to separate practice from maturational change and account for missingness and individual trajectories. The available literature supports an extended developmental program, while the relative contributions of acoustic encoding, prediction, executive control and language remain active research problems Calcus (2024) Szabó et al. (2016).
Musical Scene Analysis: Instruments, Voices, and Perceptual Fusion
Music presents an especially revealing auditory scene because physical sources may intentionally blend into a chord, split into recognizable instrumental parts or create a melody distributed across notes from different instruments. Harmonicity, common onset, coordinated modulation, timbre and musical expectations can support both fusion and segregation. Whether a listener hears one coherent musical line or several voices depends on composition, performance, listening experience and task. A straightforward count of sound-producing instruments is therefore an incomplete measure of perceived organization Bregman (1990) Siedenburg et al. (2026).
Siedenburg, Hake, Jacobsen and Bürgel's 2026 synthetic review addresses auditory scene analysis in music, connecting classic segregation principles to questions about musical streams and perceptual grouping. Its significance is to map a domain in which common acoustic causes and culturally learned musical patterns can support different organizational outcomes. Music invites experiments on orchestral mixtures, melodic continuity, source timbre, musical training and changes in attention that are difficult to capture with repeated sine tones Siedenburg et al. (2026).
Cross-domain generalization must be measured rather than presumed. A listener who separates concurrent instrumental melodies may not show the same benefit when recognizing one of several unfamiliar speakers. Musicianship can correlate with auditory skills without establishing that training causes broad improvements in everyday scene analysis. Well-controlled longitudinal, random-assignment or carefully matched comparison designs are needed to distinguish experience-dependent grouping from selection biases and task familiarity. Music serves as a powerful domain for discriminating theories of auditory grouping, not as a universal substitute for speech experiments Siedenburg et al. (2026) Kidd et al. (2005).
Evidence Synthesis: What the Experimental Traditions Establish
Several findings form the strongest convergent foundation of auditory scene analysis. Controlled manipulations of frequency separation and presentation rate change sequential streaming reports; harmonic relations and onset timing influence simultaneous grouping; coherent temporal fluctuations affect perceptual binding; ambiguous sequences permit more than one persistent organization; and listeners select behaviorally relevant speech under competition. These claims arise from distinct tasks and laboratories, and each identifies a reproducible relation between stimulus structure, context and measured behavior. They support an organized perceptual account of acoustic scenes without specifying one single universal grouping algorithm Bregman (1990) Darwin & Ciocca (1992) Elhilali et al. (2009) Winkler et al. (2012).
Human cortical recordings provide a complementary result: neural representations of acoustic features and speech structure vary with stimulation and with attentional target, linking measurable neural dynamics to perceptual competition. Evidence from cortical streaming physiology, attended-speech tracking and broader perceptual-organization networks makes it plausible that grouping and selection are distributed functions. Yet neural correlations have different causal strengths. A stimulus-driven feature response, a brain signal covarying with a report, and selective tracking of an attended voice each answer different questions about the organization process Micheyl et al. (2007) Mesgarani & Chang (2012) Zion Golumbic et al. (2013).
The limits of convergence are as informative as the successes. Speech comprehension depends on language knowledge and selective attention, while tonal streaming tasks often avoid linguistic context; texture judgments require statistical summaries, while music can promote purposeful fusion. Thus different tasks sample related but potentially nonidentical organizations. Scientific synthesis should build a layered model with interacting features, source continuity, attention and task goals, then test which component mechanisms transfer across domains. No single experiment currently supplies a complete causal map from physical mixture to perceived world Shamma et al. (2011) Szabó et al. (2016) Siedenburg et al. (2026).
Measurement Validity, Reproducibility, and Analytical Choices
A measurement of scene analysis should begin with an explicit operational question: Is the outcome source number, streaming probability, temporal-order accuracy, target intelligibility, localization, identification, percept persistence or neural tracking? These measures can dissociate even when investigators describe all of them as 'segregation.' The appropriate design includes an unambiguous physical control, an ambiguous condition and manipulations selected to differentiate target theories. Individual performance distributions matter because a group mean can obscure sharply different streaming states or attentional strategies Carlyon (2004) Winkler et al. (2012).
Reproducibility requires detailed acoustic specification: calibrated sound level, source and delivery characteristics, headphone or room configuration, sample rate, stimulus spectra, sequence timing, randomization, masker properties and analysis code. A report of harmonicity effects without characterizing peripheral excitation or onset differences may leave a rival acoustic explanation unresolved. With neural measures, additional controls should document stimulus-locked variance, participant state, statistical model selection, multiple testing, and the correspondence between trial-level perceptual reports and recorded responses Darwin & Ciocca (1992) Micheyl et al. (2007).
Experimental results are best compared when studies report confidence intervals, observer-level variability and diagnostic effects rather than only statistical significance. Repeated measures and continuous perceptual tracking require models accounting for temporal dependence; multistable durations should not be analyzed as unrelated independent trials. Hierarchical models can accommodate listener variability but should be checked for identifiability and predictive adequacy. Computational fit is valuable only when it resolves identifiable experimental contrasts, including conditions on which competing mechanisms diverge Szabó et al. (2016) Snyder & Alain (2007).
A Critical Comparison of Competing Explanatory Frameworks
Stimulus-driven grouping accounts emphasize regularities such as harmonicity, common onset and proximity. Their strength is operational precision: the experimenter can manipulate a cue and observe its effect. Their principal explanatory challenge arises when cues conflict, when training and attention change grouping, or when physically identical stimuli alternate between interpretations. Those observations require mechanisms of context, competition or state dependence to supplement a static cue assignment. The existence of these additional influences leaves the acoustic cues indispensable as inputs to scene construction Bregman (1990) Darwin (1997) Winkler et al. (2012).
Adaptation-based accounts can explain changes in sensitivity and streaming probability as sensory populations respond to repeated frequency patterns. Predictive accounts explain how temporal regularity structures candidate event descriptions, while temporal-coherence accounts emphasize synchronized variation across feature channels. Neural tracking findings add evidence that attention enhances representations of chosen signals. Each framework has a distinct measurable core, and their explanatory variables can co-occur: adaptation changes channel responses, prediction changes expectations, and attention changes the gain or use of particular features. A productive model comparison asks when these components make opposing predictions, not which vocabulary sounds most comprehensive Micheyl et al. (2007) Denham & Winkler (2006) Shamma et al. (2011).
Generative and machine-learning approaches can model source separation by inferring latent acoustic causes. Their utility depends on whether the latent variables correspond to psychological entities and whether predictions match human errors and perceptual uncertainty. An algorithm that optimizes signal reconstruction may separate components that listeners experience as one musical whole; a model that reproduces listener reports may produce acoustically inaccurate waveforms. Theoretical integration therefore calls for joint tests of reconstructive performance, reported grouping, neural response and generalization, with each outcome defined independently Szabó et al. (2016) Siedenburg et al. (2026).
Individual Variability, Context, and Generalization
Scene organization varies with peripheral sensitivity, attention, prior experience, listening goals and the acoustic context. Developmental research shows that the capacity to select speech and resolve complex auditory mixtures continues changing through childhood. Differences between individuals may likewise arise from multiple processes: reduced hearing sensitivity, distinctive frequency resolution, variation in attentional control, learned expectations or unfamiliarity with stimuli. An apparent between-group segregation difference cannot identify which process contributes without measurement of plausible covariates and convergent tasks Calcus (2024) Kidd et al. (2005).
Cultural and linguistic experience is relevant insofar as it shapes familiar rhythmic patterns, phonological expectations and musical organization. The evidence available for auditory scene analysis is heavily influenced by participants and stimuli from particular research contexts; transferring conclusions across languages, ages and listening environments is a scientific question requiring direct comparison. A musical-training association or advantage with familiar speech can result from selection effects or task-specific learning. Generalization tests should cross acoustic cue structure with linguistic and musical familiarity wherever possible Siedenburg et al. (2026) Wood & Cowan (1995).
Applied implications include listening in classrooms, workplaces, public spaces, telecommunication and assistive hearing technologies. Translational value follows from precise predictions about which acoustic configurations improve the target task under actual environmental conditions. An intervention that increases target speech intelligibility may alter spatial clarity, source continuity or listener fatigue differently; these outcomes need separate measurement. Auditory scene analysis offers a principled experimental language for such problems while clinical benefit and population-level effectiveness require purpose-designed studies Calcus (2024) Carlyon (2004).
Auditory Organization Across Biological and Artificial Systems
Artificial separation systems can estimate clean source waveforms from recorded mixtures, detect sound events and classify scene categories. These capacities make them experimentally interesting comparators for human hearing, especially when models are tested against ambiguous streams, perceptual masking and sound textures. However, the engineering objective of minimizing reconstruction error differs from the psychological task of explaining how a listener groups sounds under uncertainty. A fair comparison must fix the stimulus, define comparable behavioral or computational outputs and measure failures as well as successful identifications Szabó et al. (2016) McDermott et al. (2013).
Temporal coherence and predictive grouping provide candidate computational principles that can be instantiated in either biologically motivated or general engineering models. Empirical interpretation requires specifying representations and training conditions rather than assuming equivalence because two systems carry the same model label. A human may interpret a repetitive tone sequence as two alternating streams while an artificial separator produces two waveform tracks for an entirely different reason. A mechanism-level comparison would test response to cue conflict, learning history, attentional instructions and multistable perceptual interpretations Shamma et al. (2011) Winkler et al. (2012).
In the intellectual horizon of the Artificial Era, these comparisons address how different kinds of systems organize an acoustic environment into actionable structure. Claims of comparable auditory scene organization must rest on explicit operational criteria: invariant tracking through acoustic changes, context-sensitive grouping, uncertainty-sensitive selection and transfer across novel environments. Such empirical criteria can support comparative research without treating any one implementation as a settled psychological equivalent. The present monograph records an agenda for testable cross-system explanation, not an experimental demonstration of identity between biological and artificial perceptual mechanisms Szabó et al. (2016) Winkler et al. (2009).
Unresolved Problems and a Program of Discriminating Experiments
The strongest unresolved problem is identifying a causal architecture that links feature-level acoustic information to the perception of persistent sources. Existing approaches offer important but only partially connected answers concerning peripheral channel responses, temporal-coherence binding, prediction, attentional target selection and object persistence. Experiments should construct cue conflicts that yield different quantitative predictions for these mechanisms, then compare listener-level responses and neural measures under identical mixtures. Explicitly modeled uncertainty and state transitions are essential for ambiguous stimuli Shamma et al. (2011) Micheyl et al. (2007) Winkler et al. (2012).
A second priority is the relation between simple and naturalistic listening. Mechanisms inferred from repeating tones should predict speech and music only where the relevant stimulus dimensions and perceptual tasks match. Research should evaluate models across independent audio collections and report where they fail to explain overlapping talkers, dense sound textures, occluded sources and purposeful musical fusion. Developmental trajectories, aging, language familiarity and listener heterogeneity should be incorporated as explanatory variables rather than noise to be averaged away McDermott & Simoncelli (2011) Calcus (2024) Siedenburg et al. (2026).
A third priority is stronger triangulation of evidence. Auditory-scene studies benefit from combining well-controlled psychophysics, continuous percept reports, neural tracking, perturbation where ethically and technically justified, and comparative computational predictions. Predeclared measures, open stimuli and independent replication would allow the field to separate robust grouping regularities from context-specific task strategies. Such a program could determine where apparently competing theoretical accounts describe complementary stages and where they make genuinely incompatible claims Snyder & Alain (2007) Szabó et al. (2016) Mesgarani & Chang (2012).
Conclusion
Auditory scene analysis is the structured perceptual organization of acoustic mixtures into concurrent events, persisting auditory objects and temporally coherent streams. Its experimental foundations demonstrate that segregation depends on both physical relations among sounds and a listener's temporal context, learned expectations and attentional selection. The evidence connects frequency separation, harmonicity, onset timing, temporal coherence, continuity and competition to distinct behavioral outcomes, while neural work associates auditory and higher-order cortical dynamics with source processing and selected speech. A fundamental account therefore requires several interacting explanatory levels Bregman (1990) Darwin (1997) Micheyl et al. (2007) Shamma et al. (2011).
The critical scientific synthesis establishes a stable methodological principle: sound-source identity, objective acoustic decomposition, a reported perceptual stream and understanding a selected message must be measured separately. Their correlations are scientifically informative precisely because they do not always coincide. Alternating tones provide clean tests of temporal organization; multistability exposes underdetermination; competing voices reveal selection constraints; sound textures demonstrate statistical encoding; and music reveals circumstances where grouping and fusion are purposeful. Across these domains, controlled cue manipulations and convergent methods carry more explanatory weight than broad analogies about hearing Winkler et al. (2012) McDermott et al. (2013) Siedenburg et al. (2026).
The next advances will come from models that predict new behavioral and neural results across conflicting cues, temporal histories and naturalistic contexts, with uncertainty, individual variation and replicability explicitly represented. Such models can connect the long-established psychophysical foundations of auditory scene analysis to contemporary neuroscience, development, auditory engineering and comparative artificial perception. The field's central achievement is a method for studying how listeners recover coherent structure from complex sound; its enduring challenge is to specify the causal computations by which that structure is produced and maintained Elhilali et al. (2009) Calcus (2024) Szabó et al. (2016).
References
- Anstis, S., & Saida, S. (1985). Adaptation to auditory streaming of frequency-modulated tones. Journal of Experimental Psychology: Human Perception and Performance, 11(3), 257–271. https://doi.org/10.1037/0096-1523.11.3.257
- Bregman, A. S. (1990). Auditory scene analysis: The perceptual organization of sound. MIT Press. https://mitpress.mit.edu/9780262022972/auditory-scene-analysis/
- Calcus, A. (2024). Development of auditory scene analysis: A mini-review. Frontiers in Human Neuroscience, 18, Article 1352247. https://doi.org/10.3389/fnhum.2024.1352247
- Carlyon, R. P. (2004). How the brain separates sounds. Trends in Cognitive Sciences, 8(10), 465–471. https://doi.org/10.1016/j.tics.2004.08.008
- Carlyon, R. P., Cusack, R., Foxton, J. M., & Robertson, I. H. (2001). Effects of attention and unilateral neglect on auditory stream segregation. Journal of Experimental Psychology: Human Perception and Performance, 27(1), 115–127. https://doi.org/10.1037/0096-1523.27.1.115
- Cherry, E. C. (1953). Some experiments on the recognition of speech, with one and with two ears. Journal of the Acoustical Society of America, 25(5), 975–979. https://doi.org/10.1121/1.1907229
- Cusack, R. (2005). The intraparietal sulcus and perceptual organization. Journal of Cognitive Neuroscience, 17(4), 641–651. https://doi.org/10.1162/0898929053467541
- Darwin, C. J. (1997). Auditory grouping. Trends in Cognitive Sciences, 1(9), 327–333. https://doi.org/10.1016/S1364-6613(97)01097-8
- Darwin, C. J., & Ciocca, V. (1992). Grouping in pitch perception: Effects of onset asynchrony and ear of presentation of a mistuned component. Journal of the Acoustical Society of America, 91(6), 3381–3390. https://doi.org/10.1121/1.402828
- Denham, S. L., & Winkler, I. (2006). The role of predictive models in the formation of auditory streams. Journal of Physiology–Paris, 100(1–3), 154–170. https://doi.org/10.1016/j.jphysparis.2006.09.012
- Elhilali, M., Ma, L., Micheyl, C., Oxenham, A. J., & Shamma, S. A. (2009). Temporal coherence in the perceptual organization and cortical representation of auditory scenes. Neuron, 61(2), 317–329. https://doi.org/10.1016/j.neuron.2008.12.005
- Kidd, G., Jr., Mason, C. R., & Gallun, F. J. (2005). Combining energetic and informational masking for speech identification. Journal of the Acoustical Society of America, 118(2), 982–992. https://doi.org/10.1121/1.1953167
- McDermott, J. H., Schemitsch, M., & Simoncelli, E. P. (2013). Summary statistics in auditory perception. Nature Neuroscience, 16(4), 493–498. https://doi.org/10.1038/nn.3347
- McDermott, J. H., & Simoncelli, E. P. (2011). Sound texture perception via statistics of the auditory periphery: Evidence from sound synthesis. Neuron, 71(5), 926–940. https://doi.org/10.1016/j.neuron.2011.06.032
- McDermott, J. H., Wrobleski, D., & Oxenham, A. J. (2011). Recovering sound sources from embedded repetition. Proceedings of the National Academy of Sciences, 108(3), 1188–1193. https://doi.org/10.1073/pnas.1004765108
- Mesgarani, N., & Chang, E. F. (2012). Selective cortical representation of attended speaker in multi-talker speech perception. Nature, 485, 233–236. https://doi.org/10.1038/nature11020
- Micheyl, C., Carlyon, R. P., Gutschalk, A., Melcher, J. R., Oxenham, A. J., Rauschecker, J. P., Tian, B., & Wilson, E. C. (2007). The role of auditory cortex in the formation of auditory streams. Hearing Research, 229(1–2), 116–131. https://doi.org/10.1016/j.heares.2007.01.007
- Shamma, S. A., Elhilali, M., & Micheyl, C. (2011). Temporal coherence and attention in auditory scene analysis. Trends in Neurosciences, 34(3), 114–123. https://doi.org/10.1016/j.tins.2010.11.002
- Siedenburg, K., Hake, R., Jacobsen, S., & Bürgel, M. (2026). Auditory scene analysis in music: A synthetic review. Attention, Perception, & Psychophysics, 88(6), Article 171. https://doi.org/10.3758/s13414-026-03310-y
- Snyder, J. S., & Alain, C. (2007). Toward a neurophysiological theory of auditory stream segregation. Psychological Bulletin, 133(5), 780–799. https://doi.org/10.1037/0033-2909.133.5.780
- Szabó, B. T., Denham, S. L., & Winkler, I. (2016). Computational models of auditory scene analysis: A review. Frontiers in Neuroscience, 10, Article 524. https://doi.org/10.3389/fnins.2016.00524
- Tougas, Y., & Bregman, A. S. (1990). Auditory streaming and the continuity illusion. Perception & Psychophysics, 47(2), 121–126. https://doi.org/10.3758/BF03205976
- Winkler, I., Denham, S. L., & Nelken, I. (2009). Modeling the auditory scene: Predictive regularity representations and perceptual objects. Trends in Cognitive Sciences, 13(12), 532–540. https://doi.org/10.1016/j.tics.2009.09.003
- Winkler, I., Denham, S., Mill, R., Bőhm, T. M., & Bendixen, A. (2012). Multistability in auditory stream segregation: A predictive coding view. Philosophical Transactions of the Royal Society B: Biological Sciences, 367(1591), 1001–1012. https://doi.org/10.1098/rstb.2011.0359
- Wood, N. L., & Cowan, N. (1995). The cocktail party phenomenon revisited: Attention and memory in the classic selective listening procedure of Cherry (1953). Journal of Experimental Psychology: General, 124(3), 243–262. https://doi.org/10.1037/0096-3445.124.3.243
- Zion Golumbic, E. M., Ding, N., Bickel, S., Lakatos, P., Schevon, C. A., McKhann, G. M., Goodman, R. R., Emerson, R., Mehta, A. D., Simon, J. Z., Poeppel, D., & Schroeder, C. E. (2013). Mechanisms underlying selective neuronal tracking of attended speech at a cocktail party. Neuron, 77(5), 980–991. https://doi.org/10.1016/j.neuron.2012.12.037