Ear-voice span

Why does an interpreter who lags further behind the speaker often deliver a more faithful rendering?

Watch two simultaneous interpreters on the same speech. One is close behind the speaker, starting each phrase almost as it lands. The other is noticeably further back — long enough that a listener glancing between the booth and the podium can see the gap.

Intuition says the closer one is better: more responsive, less likely to forget, less strain on memory. But the further-back interpreter frequently produces the more faithful rendering, with fewer corrections and fewer sentences that have to be rescued halfway through. That inverts the obvious reading of the trade-off, and the inversion is what needs explaining.

The companion account of simultaneous interpreting establishes the trade-off itself: the interpreter holds a lag, that lag costs working memory, and the size of it is a tuned compromise. What that leaves open is the direction of the fidelity gradient — why, within the workable range, waiting longer buys accuracy rather than merely costing memory. That is the question here.

Start with what an interpreter is actually committing to when they open their mouth. Speech is irreversible in a way writing is not: once a word is out, it is in the listener's ears. So beginning a sentence is a commitment to a structure that the rest of the sentence must fit.

Now ask what determines whether that commitment is safe. It is safe if the speaker's meaning is already determinate at the moment of commitment. It is unsafe if the thing that fixes the meaning has not yet been said.

And that is language-dependent in a specific, predictable way. In a verb-final language, the predicate — the word that says what actually happened — arrives at the end of the clause. An interpreter working out of such a language into a verb-medial one cannot know, from the first two-thirds of the sentence, whether the subject bought, refused, considered or did not buy the thing. Everything before the verb is compatible with all of those. Commit early and you are guessing.

Similarly for negation that arrives late, for a subordinate clause that turns out to reverse the main one, and for a numeral qualified afterwards by a word meaning approximately or at most.

Now follow what a short span costs. The interpreter commits to a structure, the disambiguating word arrives and contradicts it, and they must repair — audibly. Repair is expensive twice over. It consumes the interpreter's own processing while new input keeps arriving, so it makes the next stretch worse. And it costs the listener, who has already integrated the wrong version.

A longer span avoids that by waiting for the unit of meaning to complete. The interpreter renders whole propositions rather than fragments, so each utterance is committed to only once. The lag is not idle time; it is the interval during which the input becomes determinate enough to be safe to speak.

That gives the shape: fidelity improves with lag because lag buys determinacy, and the cost of committing early is asymmetric — an early commitment that turns out wrong is far more expensive than the memory cost of having waited.

Which immediately explains why the gradient does not continue forever. Held long enough, the span exceeds what working memory can hold accurately, and the interpreter starts losing material outright — not mis-rendering it but omitting it. So the curve rises with lag, peaks, and falls. Skilled interpreters sit near the peak, and the peak sits further back for verb-final source languages than for structurally similar pairs. That last prediction is the useful one, because it says the optimum is a property of the language pair, not of the interpreter's nerve.

Anticipation is not the same as guessing, and this is the most important refinement. Skilled interpreters do commit before the disambiguating word, routinely — but they do it on the basis of genre knowledge, collocational probability, the speaker's known position, and what the conference is about. That is prediction with a high hit rate, and it is a large part of what expertise consists of. So the honest statement is not "wait for certainty" but "commit only where the prior is strong enough that repair is unlikely". A longer span is what you fall back on when the prior is weak — an unfamiliar topic, a speaker who is being deliberately careful, or a source language where the structure genuinely withholds the verb.

Fidelity is not a single quantity. Propositional accuracy improves with lag. Completeness may not — a longer span raises omission risk under load. And latency itself matters to a listener trying to intervene in a debate. Maximising accuracy alone optimises one axis of a job with at least three.

The strategies that avoid the trade-off show it is not fundamental. Interpreters working from verb-final languages start early without committing: a deliberately neutral opening, chunking into short clauses chainable in any direction, constructions keeping several continuations available. These reduce the need for lag rather than paying for it — evidence that the mechanism is structural commitment, since techniques aimed at deferring commitment are the ones that work.

What I am not claiming. Published span figures vary widely with language pair, speaker rate, density and measurement method, so I quote none. Nor do I claim measured span correlates with rated quality across interpreters; the reported relationship is not simply monotonic, and how much of expertise is span management rather than anticipation, memory or preparation is genuinely disputed.

The falsification test. If the mechanism is structural commitment, then the optimal span should be systematically longer for verb-final source languages than for structurally parallel pairs, holding speaker rate and material constant — and repairs should cluster on sentences whose disambiguating element arrived late. If span made no difference to repair rate, or repairs were distributed evenly across sentence types, the account would be wrong and the lag would be doing something else — buffering speech rate, perhaps, rather than buying determinacy.