Auditory Pathways and Sound Localization

Sound localization depends on binaural cues, differences in the signal reaching the two ears that arise purely from geometry. A sound source located to one side reaches the near ear slightly sooner and slightly louder than the far ear. The brain exploits two such cues. The interaural time difference (ITD) is the difference in arrival time of a sound at the two ears, which can be as large as about 700 microseconds for a sound directly to one side in humans (given an interaural distance of about 22 cm) and zero for a sound directly ahead or behind. The interaural intensity difference, also called interaural level difference (ILD), is the difference in sound pressure level at the two ears, created because the head casts an acoustic "shadow" that attenuates high-frequency sound more than low-frequency sound. These two cues are complementary rather than redundant, a relationship formalized as the duplex theory of sound localization (Lord Rayleigh, early twentieth century). Low-frequency sounds (below roughly 1.5 kHz) have wavelengths longer than the head's diameter, so they diffract around it with little intensity loss — but their long, slowly-varying waveform allows the nervous system to track phase differences between the ears reliably, making ITDs the dominant cue. High-frequency sounds (above roughly 2 kHz) have wavelengths shorter than the head, so they cast a genuine acoustic shadow producing large ILDs, but their rapid oscillation makes the auditory nerve unable to phase-lock to individual cycles precisely enough for ITD coding to remain unambiguous. Between about 1.5 and 2 kHz lies an ambiguous zone where neither cue works cleanly, which is one reason human sound localization is measurably poorer for tones in that narrow band. Both cues require binaural convergence, meaning that afferents from the two cochleae must meet on individual neurons for a comparison to occur. This first happens in the superior olivary complex of the brainstem, making it the first true site of spatial hearing in the ascending pathway. Everything downstream — the inferior colliculus, the medial geniculate nucleus of the thalamus, and auditory cortex — inherits and refines this binaural computation while preserving the tonotopic (frequency-mapped) organization established in the cochlea, so that the ascending auditory system simultaneously carries a "what" code (frequency) and a "where" code (location) in parallel.

Step 1 — From hair cell to auditory nerve. Sound-evoked vibration of the basilar membrane deflects stereocilia on inner hair cells, which release glutamate onto spiral ganglion neuron endings in proportion to instantaneous membrane displacement. Critically, for frequencies below about 3–4 kHz, auditory nerve fibers phase-lock: they fire preferentially at one point in each cycle of the stimulus waveform, so the pattern of spike timing across many trials/fibers statistically encodes the fine temporal structure of the sound, not merely its average rate. This phase-locked timing code is the raw material ITD detection depends on.

Step 2 — Cochlear nucleus: the first synapse, and multiple parallel streams. Every auditory nerve fiber bifurcates upon entering the brainstem, sending one branch to the dorsal cochlear nucleus (DCN) and one to the ventral cochlear nucleus (VCN). This is the obligatory first synapse for all auditory information — there is no ascending auditory information that bypasses the cochlear nucleus. The VCN contains bushy cells, whose large, secure "endbulb of Held" synapses preserve the auditory nerve's precise timing with minimal jitter, making them ideal relays for binaural timing computations. VCN stellate cells and DCN neurons instead extract spectral and intensity features and project more diffusely, including to the inferior colliculus directly, contributing to monaural spectral cues (important for elevation judgments) alongside the binaural pathway.

Step 3 — Superior olivary complex: where binaural comparison first happens. Bushy cell axons cross and diverge to reach both left and right superior olives, so each side of the brainstem receives excitatory input from both ears — the essential prerequisite for any binaural computation.

• In the medial superior olive (MSO), principal neurons receive separate excitatory glutamatergic inputs from the ipsilateral and contralateral cochlear nuclei on their two dendritic arbors, plus inhibitory glycinergic input. Each MSO neuron acts as a coincidence detector: it fires maximally when the excitatory postsynaptic potentials from the two ears arrive simultaneously at the cell body, which happens only for sounds with a specific ITD. Different neurons have systematically different internal axonal conduction delays (via differing axon lengths from each cochlear nucleus), so that across the population of MSO neurons, every behaviorally relevant ITD is represented by a distinct subset of "best-tuned" cells. This is the Jeffress delay-line model.
• In the lateral superior olive (LSO), principal neurons receive excitation from the ipsilateral ear and inhibition (via the medial nucleus of the trapezoid body, MNTB, which inverts an excitatory contralateral signal into a fast glycinergic inhibitory one) from the contralateral ear. LSO neurons therefore compute an excitatory-minus-inhibitory subtraction: a sound louder in the ipsilateral ear produces more firing, a sound louder in the contralateral ear silences the cell. This makes the LSO the primary site of ILD computation, and it is dominated by high-frequency-tuned neurons, consistent with the duplex theory.

Step 4 — Inferior colliculus (IC): the obligatory midbrain relay. Essentially all ascending auditory pathways, from both MSO and LSO as well as direct cochlear nucleus and DCN projections, converge on the central nucleus of the inferior colliculus. Here ITD and ILD information — computed by physically separate circuits — is combined for the first time into more unified spatial-receptive-field-like responses, and monaural spectral cues from the DCN are integrated as well (important for distinguishing front from back and judging elevation, since these cannot be solved by ITD/ILD comparison alone). The IC also imposes strong tonotopic layering: it is arranged in tonotopic laminae, frequency-organized sheets of neurons that receive convergent input from frequency-matched regions of the ascending nuclei.

Step 5 — Medial geniculate nucleus (MGN) of the thalamus. The IC projects to the ventral division of the MGN, which relays the tonotopically organized, spatially informed signal to cortex; as in other sensory thalamic relays, this stage also imposes attentional and state-dependent gain control via corticothalamic feedback.

Step 6 — Primary auditory cortex (A1) and beyond: the tonotopic map. MGN axons terminate in primary auditory cortex, located on the superior temporal gyrus (within Heschl's gyrus in humans), arranged so that neurons responding best to similar frequencies lie in continuous isofrequency bands — the cortical tonotopic map, first mapped electrophysiologically in cats and monkeys and now routinely visualized in humans with fMRI and MEG, running in an orderly gradient from low to high best-frequency across the cortical surface. Overlaid on this frequency map are more distributed, less strictly topographic representations of ITD and ILD; unlike vision's crisp retinotopic map of space, auditory cortex in mammals does not show an unambiguous point-for-point spatial map, and location appears to be represented more by broadly tuned, overlapping populations (a "meter" or hemifield code) than a Jeffress-like place code. This foreshadows the more general problem of how cortical maps represent continuous variables — sharply, as in tonotopy, or by population codes, as appears to be the case for auditory space in mammalian cortex.

Lord Rayleigh and the duplex theory (1875–1907). John William Strutt (Lord Rayleigh) tested human localization of pure tones and diffuse noise and noticed a striking discrepancy: listeners localized low tones well using timing-like cues but struggled with mid-frequency pure tones, while high-frequency tones were localized using apparent loudness differences. He worked out mathematically that a rigid sphere the size of a human head would produce negligible intensity differences for long wavelength (low-frequency) sound, since it diffracts around the head, but substantial shadowing for short-wavelength (high-frequency) sound. This established the duplex theory: ITD for low frequencies, ILD for high frequencies, with a well-documented "crossover" region of poor localization near 1.5–2 kHz where neither cue is fully effective. Later psychoacoustic work (Stevens and Newman, 1936, using tones presented outdoors on a rooftop to minimize echoes) confirmed this crossover pattern directly with human localization-error curves that rose in the mid-frequency range and fell at both extremes.

Lloyd Jeffress and the coincidence-detector model (1948). Jeffress proposed, on purely theoretical/anatomical grounds, that the nervous system could measure ITDs using an array of neurons each receiving input from both ears via axons of systematically different lengths, so each neuron would respond maximally only when a specific external ITD exactly compensated for its particular internal conduction-delay mismatch — a "place" code for time. At the time there was no direct physiological confirmation; it was an elegant hypothesis awaiting a testable circuit.

Konishi and colleagues in the barn owl (1980s–1990s). Masakazu Konishi's laboratory chose the barn owl (Tyto alba) because it hunts in total darkness using hearing alone and has an unusually precise, experimentally tractable sound-localization system with an actual computed map of auditory space in its midbrain (the external nucleus of the inferior colliculus), a feature not clearly present in mammalian cortex. Using single-unit recordings, Konishi, Terry Takahashi, Catherine Carr, and colleagues showed:

• Neurons in the owl's nucleus laminaris (the avian homolog of the mammalian MSO) are genuine coincidence detectors, firing maximally to specific ITDs regardless of the sound's overall intensity or exact frequency — precisely as Jeffress's model predicted.
• These neurons are arrayed systematically, and anatomical tracing (Carr and Konishi, 1990) directly visualized the predicted delay lines: axons from the two ears entering nucleus laminaris with graded, systematically increasing conduction delays, physically confirming the Jeffress circuit rather than merely inferring it from response properties.
• Downstream, in the owl's external nucleus of the inferior colliculus, ITD and ILD information converge onto neurons with genuinely space-specific receptive fields, forming an explicit map of auditory space — each locus in the nucleus responds best to sound from one specific direction, verified by systematically moving a sound source around the owl while recording single units (Knudsen and Konishi, 1978).
• Behaviorally, owls with one ear plugged, or fitted with prism-like devices that shift their visual field, showed predictable, quantifiable errors in orienting/striking behavior toward sound sources, and — critically — young owls raised with an optically shifted visual world recalibrated their auditory space map to match the shifted vision (Knudsen and Knudsen, 1989), demonstrating that the ITD/ILD map is experience-dependent and can be shaped by visual feedback during a sensitive period.

Mammalian confirmation and refinement. Direct intracellular and in vivo recordings in gerbils and cats (Yin and Chan, 1990; Joris, Smith, and Yin, 1998) confirmed MSO neurons behave as coincidence detectors with best-ITDs matching behaviorally relevant ranges, and demonstrated LSO neurons perform excitatory-inhibitory subtraction for ILD exactly as the anatomical wiring predicts. However, more recent work (Grothe and colleagues, 2000s) revealed an important refinement: unlike the owl, mammalian MSO neurons' best-ITDs may not be as evenly and symmetrically distributed across the full physiological range as originally assumed by a strict Jeffress model; instead, many mammalian species appear to rely more on the slopes of broadly tuned ITD functions and a comparison between the two hemispheric populations (a "two-channel" or hemifield code) rather than a precise labeled-line place code. This shows the Jeffress model is correct in its core coincidence-detection logic but was refined, not overturned, by later comparative physiology — a case study in how a good model gets sharpened rather than discarded.

Example 1 — Calculating an ITD and predicting localizability. A sound source is positioned 30° to the right of the midline. Assume the speed of sound is 343 m/s, the head can be approximated as a sphere with an effective interaural path difference described by the Woodworth formula: ITD ≈ (r/c) × (θ + sin θ) where r is head radius (≈ 8.75 cm, giving an average adult interaural distance of ~17.5 cm point-to-point, though acoustic path length around the head is somewhat longer), c is the speed of sound, and θ is the azimuth in radians. Step 1: Convert 30° to radians: θ = 0.524 rad. Step 2: θ + sin θ = 0.524 + 0.500 = 1.024. Step 3: r/c = 0.0875 m / 343 m/s = 2.55 × 10⁻⁴ s = 255 μs. Step 4: ITD ≈ 255 μs × 1.024 ≈ 261 μs. This falls comfortably within the physiological range MSO neurons can encode (roughly 0–700 μs in humans, corresponding to azimuths from 0° to 90°). Now ask: could a listener localize a 4 kHz pure tone at this position using ITD alone? A 4 kHz tone has a period of 250 μs — shorter than the computed ITD of 261 μs. This means the interaural phase difference is ambiguous (it could equally represent a delay of 261 μs or 261 μs minus a whole period), so phase-locking based ITD cues become unreliable above about 1.5 kHz, exactly the frequency limit predicted by the duplex theory, and this listener would need to rely on ILD instead. This demonstrates why the duplex theory is not arbitrary — it's a direct consequence of the geometric relationship between wavelength, head size, and neural phase-locking limits.

Example 2 — A clinical localization deficit: unilateral conductive vs. sensorineural loss. A patient reports difficulty telling which side a sound comes from, and audiometry shows a 40 dB conductive hearing loss in the left ear (e.g., from otosclerosis) with normal right-ear thresholds and normal bone conduction. Reasoning through the mechanism:

• A conductive loss attenuates sound reaching the cochlea on the affected side but does not distort neural timing once sound does get through (bone conduction thresholds are normal, meaning the cochlea and nerve are intact).
• This creates an artificial, constant ILD bias: any sound will now seem louder in the right ear regardless of true source location, because the left ear's input is uniformly attenuated by ~40 dB before it even reaches the hair cells.
• Consequence: the LSO circuit, which computes ipsilateral-excitation-minus-contralateral-inhibition, will be systematically biased — sounds actually located slightly left of midline may still produce a net "right ear louder" signal, causing the patient to mislocalize sounds toward the right (the unaffected side).
• Now compare this to a hypothetical unilateral auditory neuropathy (a sensorineural timing disorder where the cochlea transduces normally but auditory nerve firing becomes desynchronized/poorly phase-locked). Loudness (ILD) would be roughly preserved but ITD-based low-frequency localization would be selectively devastated, because coincidence detection in the MSO depends on precisely phase-locked spike timing, not just spike rate.
• This dissociation — a conductive/level-based deficit degrading ILD-dependent (mainly high-frequency) localization versus a temporal/neuropathic deficit degrading ITD-dependent (mainly low-frequency) localization — is a direct clinical extension of the duplex theory and is used diagnostically: localization testing across frequency bands can help distinguish where in the pathway (external/middle ear conduction vs. cochlear nerve timing) a lesion lies.

Masakazu Konishi built his career on a wager that a nocturnal predator would reveal principles too subtle to see in laboratory animals evolved for vision-dominated behavior. Trained originally in ethology under Peter Marler (studying bird song learning), Konishi shifted much of his later career to auditory neuroscience specifically because the barn owl offered something almost no mammal does cleanly: a behaviorally essential, purely acoustic localization task (striking prey in total darkness) paired with an genuinely explicit neural map of space that could be probed with single-unit recording, lesions, and developmental manipulation. His laboratory's decades-long dissection of the owl's ITD and ILD pathways, capped by direct anatomical visualization of Jeffress-style delay lines, is often cited as one of the cleanest examples in systems neuroscience of a computational model proposed on purely logical grounds later being confirmed circuit-by-circuit in real neurons.

Hearing aids and cochlear implants. Modern hearing aid algorithms explicitly model ILD and, increasingly, attempt to preserve fine timing cues in bilateral fittings, because patients fit with only one aid or implant lose access to binaural cues and show measurable localization deficits; bilateral cochlear implantation is now pursued partly to restore rudimentary ITD/ILD processing, though current devices still transmit envelope information poorly enough that ITD sensitivity remains far coarser than in normal hearing.

Diagnosing retrocochlear and brainstem pathology. Because ITD and ILD processing depend on specific, anatomically separable brainstem nuclei (MSO, LSO, trapezoid body), tests of binaural processing (e.g., masking-level difference tests, dichotic listening) are used clinically to localize brainstem lesions, auditory neuropathy, and certain demyelinating conditions like multiple sclerosis that disrupt fast, precisely timed conduction.

Auditory prosthetics and sensory substitution devices. Engineers designing directional microphones, robotic sound-localization systems, and assistive listening devices for the blind directly borrow the Jeffress coincidence-detection algorithm and the ILD-subtraction principle as computational strategies, since they remain among the most efficient known solutions to real-time sound-source localization.

Bioacoustics and conservation. Passive acoustic monitoring of endangered species (whales, owls, songbirds) uses microphone arrays and cross-correlation algorithms mathematically equivalent to ITD coincidence detection to localize calling animals across large distances, informing population surveys and anti-poaching efforts without requiring visual sighting.

Basic neuroscience of cortical maps. The auditory system's clean separation of a preserved, low-level tonotopic map from a more distributed, computed spatial representation is a widely used teaching and research model for how the brain balances explicit topographic coding against population coding — a theme that recurs across sensory systems and directly foreshadows broader questions in Sensory Maps and Cortical Topography.

Ask the tutor: How would damage limited to just the lateral superior olive change a person's ability to localize a low-pitched voice versus a high-pitched whistle?