Two different levers on the same process
Audiovisual integration is the brain's habit of merging a sight and a sound into one perceived event. Two very different things can push that merging around. One is timing: how far apart the sight and sound arrive. The other is meaning: whether the sound fits the picture, as when a bark accompanies a dog versus a meow. Studies that manipulate meaning and studies that manipulate timing tend to measure different things, which is the first clue that they are not interchangeable [1][3][4].
The timing work is clearest about what it changes. When people are trained on the sound-induced flash illusion, a classic trick in which one flash paired with two beeps is seen as two flashes, their judgments of simultaneity become sharper: the point at which they feel sight and sound coincide moves closer to true zero, and the window of acceptable delay narrows [4]. That window, often called the temporal binding window, is the range of lags within which the brain still treats two signals as one event. Training tightens it, which is a change in the low-level timing machinery itself [4][5].
Meaning changes what survives distraction
Semantic association seems to act at a different point. In one study, participants watched animal pictures while hearing sounds that were congruent, incongruent, or unrelated, all while under no, low, or high attentional load and while ignoring the sounds. Higher load did not break the integration of congruent picture-sound pairs, but it did suppress the extra alerting pull of incongruent or unrelated sounds [1]. In other words, meaning determined which pairings held up when attention was stretched, rather than changing the basic timing window.
That is a useful contrast with the timing studies. Timing training alters how close two signals must be to fuse [4][5]; semantic congruence alters whether a pairing is integrated robustly enough to withstand a busy attentional state [1]. The two effects can coexist because they are doing different jobs: one sets the allowable delay, the other weights how much a given pairing matters.
Meaning can color timing judgments without replacing them
Meaning is not entirely separate from timing, though. When researchers varied the semantic congruency and emotional valence of non-speech audiovisual stimuli, congruency affected both how fast people responded and how well they could discriminate the objective order of the sights and sounds, and valence had its own effect [3]. That means a higher-level property of the pairing can shift performance on a timing task, which is exactly the kind of overlap that makes the two factors hard to fully disentangle.
The most direct attempt to look at both at once used brain responses in a modified oddball task, comparing semantically meaningful audiovisual contexts and tracking the temporal dynamics of integration across early time windows [2]. Because that work is reported only in summary form, it is best read as consistent with meaning and timing both contributing, rather than as settling how they divide the labor.
So which one should you think of as primary?
Neither, on the current evidence. Timing sets the boundary conditions: signals too far apart will not be bound, and training can move that boundary [4][5]. Meaning then operates on top of whatever binding occurs, shaping how strongly a pairing is integrated and how it is judged, especially when attention is divided [1][3]. A reasonable working picture is a timing gate that decides whether fusion is possible, plus a meaning-based weighting that decides how much the fused event matters and how it is evaluated.
One caveat sits beside the semantic findings rather than undercutting them: the load study tested animal pictures with congruent, incongruent, or unrelated sounds under selective attention, so its conclusion about meaning protecting integration is scoped to that kind of pairing and task [1]. The timing findings come largely from illusion training and simultaneity judgments, which capture the binding window rather than semantic evaluation [4][5]. The two literatures are measuring different faces of integration, which is why they can both be right without one explaining the other away.
Sources used in this answer
Whether attentional loads influence audiovisual integration depends on semantic associations
Under modality-specific selective attention, higher attentional load did not disrupt integration of semantically congruent audiovisual stimuli but suppressed alerting effects from incongruent or unrelated sounds, indicating semantic association modulates load effects on integration.
Investigating Cross-Modal Temporal Dynamics and Semantic Incongruency in Multisensory Integration
Using ERP difference waves in a modified oddball paradigm, the study examined temporal dynamics of audiovisual integration in semantically meaningful contexts, addressing both semantic congruency and timing.
Evaluating the effect of semantic congruency and valence on multisensory integration
Semantic congruency and valence of non-speech audiovisual stimuli affected response times and the ability to discriminate objective temporal order.
Audiovisual illusion training improves multisensory temporal integration
Five days of sound-induced flash illusion training shifted the point of subjective simultaneity toward zero and narrowed the temporal binding window, improving audiovisual temporal integration precision.
Perceptual training narrows the temporal binding window of audiovisual integration in both younger and older adults
Audiovisual temporal discrimination training narrowed the temporal binding window in both younger and older adults, though the broader reduction in crossmodal interactions seen in younger adults did not appear in older adults.