6.3 ITDs in the MSO: the Jeffress coincidence detector
In 1948, Lloyd Jeffress proposed a model for how a neural circuit could measure interaural time differences. The model has dominated the field ever since, though modern measurements have complicated it.
Picture two axons, one carrying spike trains from the left AVCN bushy cells (representing left-cochlea input), the other from the right. Each axon extends along a row of postsynaptic neurons — the coincidence detectors — and the axons run in opposite directions. The left axon enters from the left and propagates rightward; the right axon enters from the right and propagates leftward.
Each coincidence detector along the row receives, at any moment, the left axon’s signal after a delay corresponding to the detector’s position from the left, and the right axon’s signal after the complementary delay. The detector fires when the two inputs arrive simultaneously — when the left and right spike trains coincide.
Now suppose a sound arrives at the right ear before the left. The right cochlear-nucleus axon fires first; its spike begins propagating leftward along the row. The left cochlear-nucleus axon fires slightly later (after the ITD has elapsed); its spike begins propagating rightward. Where do the two spikes meet? At the position whose internal delay (from the right end) equals the external delay (the ITD). That detector fires. The other detectors do not.
The position along the row therefore encodes the source’s azimuth. The MSO is a place code for space — a one-dimensional axis along which different positions represent different ITDs.
The interactive below renders this.
0 μs
drag ITD to move the source in space; the detector with matching internal delay fires loudest
The Jeffress model has stood the test of time in remarkable ways. The barn owl’s nucleus laminaris (the avian analog of the mammalian MSO) is a textbook delay-line structure, with anatomically and physiologically measured delays of just the kind Jeffress predicted. In mammals, the situation is more nuanced: the MSO does use coincidence detection, but the delay lines are short and the precise mechanism of how internal delays are generated is still debated (Joris & Yin and colleagues have argued that internal-difference-of-latency coding may complement or replace pure delay-line architecture). What is uncontroversial is that the MSO neuron is the place where the brain first computes a spatial code from interaural timing.
The physics demanded of this computation is breathtaking. To resolve a 10 μs ITD, the postsynaptic neuron must integrate inputs from both ears with sub-millisecond timing precision and produce a reliable spike-rate change. This requires the membrane time constant to be unusually short (a few hundred microseconds, achieved through specialized potassium channels), the dendrites to have low input impedance (to summate inputs without filtering them too heavily), and the synapses to be temporally precise (which is why we needed those endbulbs at the previous stage). The MSO is biologically one of the most heavily engineered neural circuits in the body.
⏳The history— Jeffress and the delay-line model
In 1948 Lloyd Jeffress proposed a neural circuit for measuring interaural time differences: an array of coincidence-detector neurons, each receiving input from both ears through axons of different lengths. A sound arriving at the left ear first would travel further along the left delay line before reaching the coincidence detector that compensates for exactly that ITD. The detector fires maximally when the two signals arrive simultaneously — converting a time difference into a place of maximum activity.
The Jeffress model became the canonical picture of binaural processing for half a century. In birds, the nucleus laminaris implements something very close to it, with anatomically measurable delay lines. In mammals, the medial superior olive (MSO) performs ITD computation, though the mechanism appears to rely more on inhibitory timing than on pure axonal delay. The model’s enduring value is conceptual: it shows how a temporal code can be converted to a place code using nothing more than conduction delays and coincidence detection.