👂 Cocktail Party Effect

In 1953, Cherry played listeners two spoken messages at once — through headphones, one message per ear, or both mixed into a single ear — and had them "shadow" (repeat back) only one message in real time. Listeners could do this fluently for the attended message but retained almost no information about the unattended one, coining the term "cocktail party effect" for our ability to focus on one voice among many.

Each ear receives a slightly different version of a sound depending on where it originates: interaural time differences (ITDs) — the sound arrives at the near ear microseconds before the far ear — and interaural level differences (ILDs) — the far ear is in the "sound shadow" of the head, especially for higher frequencies. The brainstem compares these cues across both ears to estimate direction, and streams from different directions are much easier to segregate perceptually, which is why the six talkers here are placed at different angles and panned in 3D rather than all coming from the same point.

Auditory scene analysis, the framework developed by Albert Bregman, shows that the brain also groups sound energy into streams using non-spatial cues — especially fundamental pitch (voice register) and rhythm/timing. Two voices from the same direction can still often be told apart if one is higher-pitched or has a different speaking cadence. This simulator gives each synthesized talker both a distinct base pitch and a distinct syllable-like rhythm pattern, so segregation works even before spatial cues are considered.

With the Attention Strength slider at 0, all active talkers play at equal gain — nothing is filtered, modeling divided attention spread thinly across every voice (and in real life, comprehension of any single one suffers). As you raise the slider, the focused talker is progressively boosted and every other talker is progressively attenuated, modeling increasingly strong selective attention toward a single conversational stream, at the cost of picking up almost nothing from the others.

Neville Moray's 1959 follow-up to Cherry's work found that listeners almost never recalled ordinary unattended content, but roughly a third of the time they did notice their own name spoken in the unattended channel. This shows attentional filtering is not absolute — some background processing of unattended input continues, and sufficiently salient or self-relevant content can interrupt and redirect attention. The blip here is deliberately given a distinctive rising three-note contour and is not run through the same attenuation as ordinary background speech, so it should catch your ear regardless of where your focus slider is set.

It is an estimated signal-to-noise ratio in decibels for the focused talker relative to the combined loudness of every other active talker, computed as 20·log10(amplitude of focused talker ÷ RMS amplitude of the rest). At Attention Strength 0 this hovers near 0 dB (all voices roughly equally loud); as you increase the strength toward 1, the focused talker's effective SNR climbs well above the others, mirroring how selective attention functionally improves the "signal" you can extract from one voice in a noisy crowd.

Each talker is routed through a Web Audio PannerNode using HRTF (head-related transfer function) panning, which simulates how a sound's spectrum and interaural timing change with direction as if it were bouncing off your own head and ears. Loudspeakers let both ears hear both channels and introduce room reflections that blur this effect, so headphones give a much more convincing sense that each talker is coming from its plotted position around you.

No — like the rest of this site, it uses no external audio files. Each "voice" is synthesized live from oscillators (sawtooth, triangle or square waves) with slight pitch vibrato and a formant-shaping filter, gated by a scheduled envelope that produces short syllable-like bursts in a fixed rhythm pattern unique to that talker. It is not intelligible speech, but it reliably produces distinguishable, simultaneous "voices" that differ in pitch, timbre and cadence — enough to demonstrate genuine auditory stream segregation without using any pre-recorded audio.