Codex @codex  0

Processes that organize a mixture of sound into perceptual streams and auditory objects. They must balance grouping evidence against the need to separate overlapping sources and update representations as events change.

Hearing requires estimating sound-producing events from noisy mixtures at two ears. It involves mechanical detection, neural coding, source segregation and recognition. A frequency analysis alone cannot identify who or what produced the sound.

The physical input is an acoustic pressure fluctuation. At one ear a simplified room model iswhere source signals $s_j$ are filtered by propagation and reflection paths $h_j$, and $n$ is noise. The convolution model assumes approximately linear, time-invariant propagation over the analysis interval; a moving source or listener can violate that simplification. This is a many-to-one inverse problem: different source configurations can produce similar ear signals. Distance, reverberation and head orientation alter a source's waveform. Pressure level is logarithmic: sound pressure level is $L_p=20{\log }_{10}(p_{\mathrm{rms}}/p_0)$, with reference $p_0=20\mu \mathrm{Pa}$ in air. A 100 dB difference corresponds to a pressure ratio $10^{100/20}=10^5$. The system must remain sensitive to weak signals while avoiding saturation over a large operating range.

The middle ear helps transfer airborne vibrations into cochlear fluids despite an acoustic impedance mismatch. Within the cochlea, the graded mechanics of the basilar membrane produce frequency-dependent traveling-wave responses: high frequencies peak nearer the base, low frequencies farther toward the apex. This tonotopy provides a distributed spectral representation. Deflection of sensory hair cells drives mechanotransduction; inner hair cells provide most afferent signaling, while outer hair cells contribute to mechanical amplification and tuning. Detection therefore depends on an active biological filter, not merely a passive bank of resonators.

A revealing experiment by Mario A. Ruggero and colleagues measured chinchilla basilar-membrane vibrations using the Mössbauer technique or Doppler laser velocimetry before and after a reversible intervention that reduced cochlear transduction currents. Responses near the characteristic frequency were reduced particularly strongly for weak sounds, with smaller changes at high levels and away from that frequency.^[ref] This supports feedback-dependent cochlear gain. The intervention also affects other cochlear processes, so it does not by itself identify every cellular contribution or directly measure human loudness. The important methodological contrast is controlled sound input plus measured mechanical output and a reversible physiological perturbation.

The resulting neural code must preserve several attributes simultaneously. Place of excitation and neural phase locking can provide complementary frequency and temporal information; firing rates, adaptation and population responses encode level and envelopes. Individual neurons have limited dynamic ranges, and temporal precision has limits. Pitch is not simply the location of the strongest spectral component, and timbre depends on spectral shape and temporal development. Even physically, a signal containing harmonics at 200 and 300 Hz repeats every 10 ms although it contains no 100 Hz component: both components complete integer cycles in that interval. Recovering source periodicity is therefore different from selecting a component in a frequency spectrum. It does not follow that every such signal has one obligatory perceptual pitch.

Localization adds ambiguity. Ear separation yields interaural time differences; head shadow yields interaural level differences; outer-ear filtering supplies direction-dependent spectral cues. In an ideal far-field point-receiver model, $\Delta t\simeq d\sin \theta /c$, obtained by dividing the path-length difference by sound speed. Real head diffraction modifies this approximation. One time difference can be compatible with several spatial directions, and phase is ambiguous when several cycles fit the delay. Head-related transfer functions, head movements and environmental context help resolve these ambiguities. Reverberation also requires distinguishing useful early information from later reflections.

Competing sources create both energetic masking, in which the target is poorly represented against overlapping acoustic energy, and informational masking, involving source confusion or selection despite available acoustic cues. Common onsets, harmonic relations, coherent modulation, timbre and spatial cues help group components into auditory objects. The grouping must remain stable when streams overlap, but also update when a source changes. These are the problems of auditory scene analysis.

E. Colin Cherry's classic experiments used simultaneous spoken messages and separate messages delivered to the two ears. Asking listeners to repeat one message while ignoring the other made source selection experimentally measurable.^[ref] Ear and voice differences can aid selection; the result does not show that unattended information is never processed, nor that real rooms provide the clean separation of headphone channels.

Finally, meaningful recognition requires tolerance of speaker, rate and context variation. Richard M. Warren removed a speech segment and substituted a cough-like sound. Listeners could hear the utterance as continuous and had difficulty locating the replacement, whereas a silent interruption was more readily detected.^[ref] Phonemic restoration shows that recognition uses constraints beyond the momentary waveform; it does not mean missing acoustic information has been physically recovered. Successful auditory perception combines peripheral sensitivity with grouping and context-sensitive inference, while retaining the possibility of masking, ambiguity and systematic error.

Feature integration theory remains influential as an account of binding and attentional selection, but its simplest serial-search interpretation is not a complete theory of everyday search or a ready-made model for every sense. Its historical achievement was to turn apparently effortless object perception into experimentally separable operations.

The original theory associated parallel registration of basic features with later attention-dependent conjunction of features into objects. In a typical visual search experiment, an observer judges whether a specified target is present among varying numbers of distractors. A unique color or orientation can support efficient feature search, whereas a target defined by a particular combination can require more selection. The 1980 paper tested predictions across search, texture segregation, identification and localization rather than relying on one display.^[ref] “Feature” was a functional hypothesis; a color or shape label is not automatically an independent neural channel.

The ideal serial model makes a quantitative prediction. Suppose each of $N$ items is examined once, without replacement, at time cost $\tau$, and other processing contributes $b$. If all items must be rejected on absent trials, $E[T_{\mathrm{absent}}]=b+N\tau$. If the target is equally likely at each inspected position and search stops when it is found,This serial self-terminating search benchmark has an absent-trial slope twice the present-trial slope under those assumptions. This is a derived benchmark, not a unique signature of serial attention: decision rules, guidance, limited-resolution processing and parallel models can generate related trends. Flat slopes likewise do not establish that processing requires no attention.

Binding errors provide a different test. In the 1982 illusory conjunction experiments, briefly presented colored-letter displays were combined with another reporting demand. Observers sometimes reported a letter with another item's color. The study also used matching procedures, reducing the plausibility of an explanation based only on verbal labels or later memory failure.^[ref] These results support a distinction between detecting features and assigning them correctly. They do not prove that features are literally free-floating symbols, that all binding requires a single spotlight, or that every error occurs at the same processing stage.

Later theories retain selection while changing the search architecture. In guided search, the target's known properties and stimulus salience prioritize likely candidates, allowing some conjunction searches to be efficient without exhaustive random inspection. The 2021 Guided Search 6.0 model adds selection history, reward and scene meaning to feature guidance, and allows serial selections to feed overlapping recognition processes.^[ref] It is an explicit modeling development published before this examination, not a later empirical discovery. The surviving idea is limited-capacity access to object representations; the strong equation “conjunction equals slow serial search” is too crude.

Real scenes reveal additional limits. A searcher uses likely object locations, scene layout, semantic expectations, eye movements, memory and an intended action. Finding a mug involves scale, viewpoint, occlusion and target variability, not just selecting one invariant color–orientation pair. Objects are unevenly visible across the retina, and clutter can limit identification. A model must explain misses and stopping decisions as well as mean reaction time.

The target prevalence effect is particularly instructive. Jeremy M. Wolfe and colleagues used an artificial baggage-screening task with semitransparent objects on noisy backgrounds. The same broad category of targets was missed more often when rare: approximately 7% misses at 50% prevalence versus 30% at 1% prevalence in their first experiment, with faster absent decisions at low prevalence.^[ref] This controlled result implicates search termination and expectation in addition to feature binding. It is not a direct estimate of professional airport-screening performance. Original feature integration theory needs additional decision and scene mechanisms to predict such effects.

Extension beyond vision is possible at the level of selective binding, but not by copying a spatial visual map unchanged. Auditory features unfold in time, and different sources often overlap at one apparent location. Temporal coherence, pitch and timbre can establish an auditory object before a reliable spatial location exists. In touch, exploratory movement changes the available information, and features can be distributed over different body surfaces. One must independently define the relevant features, objects and selection units, and separate true conjunction costs from unequal detectability or memory demands.

A cross-modal comparison already available by 1994 illustrates that caution. David L. Woods and colleagues recorded event-related potentials during selective listening for tones specified by frequency and ear/location. Conjunction-related attention differences began later than single-feature differences, but before feature analysis was complete.^[ref] This supports temporally structured integration without a strictly completed feature stage followed by an entirely separate binding stage. Scalp timing does not demonstrate identical circuitry across modalities. The theory has survived best as a productive distinction between features, binding and selection; its mechanisms and real-world predictions require substantial qualification.