Should You Speak or Type?
Many of us speak our instructions into a smartphone while on the move, then switch to typing at a keyboard when we're at a desk handing tasks to AI. But you may have noticed that even when you give the exact same instructions, the quality of the output comes out subtly different. That feeling is probably even more familiar now that dictating notes and instructions by voice has become part of daily life.
It might not just be in your head. A paper posted to arXiv by researchers led by Zizhao Hu shows experimentally that different input channels leave behind different kinds of noise — and that this difference actually splits LLM answer quality. Typing leaves behind typos, while speech leaves traces of transcription disfluency and sentence restructuring, and the burden these place on AI turns out to be different.
What the Study Did
The researchers built a toolkit called HIVE (Human Input-Variation Engine). It systematically simulates the typo patterns that occur when people type on a QWERTY keyboard, as well as the transcription variations that arise when speech is converted to text. The speech-side variation comes in two flavors: the filler words and hesitations left behind by ordinary transcription tools, and the smoothing that AI dictation tools perform when they reorganize spoken language into polished sentences. The team fed these altered prompts into a range of LLMs and measured how accuracy changed. The strength of the study lies in comparing, under matched conditions, the traces left by both input methods side by side.
What They Found
First, speech-transcription variation lowered accuracy across every instruction-tuned model tested. The cause is interesting: the cost came not from filler words, but from changes to sentence structure. Second, the hit from keyboard typos was far smaller — models absorbed a substantial amount of typos before accuracy began to break down.
Third, both phenomena traced back to a single root cause: how many of the question's original tokens survived. Variations that destroyed tokens were harmful, while simply appending new tokens alongside the original ones cost almost nothing. Because sentence restructuring erases the original phrasing itself, it does more damage to voice-input accuracy.
Fourth, the gap between the two channels only showed up in tasks that required constructing an answer or reasoning toward one — it disappeared on multiple-choice tasks. In other words, for tasks where you just have to pick an answer, the input channel doesn't matter; but for tasks where you have to generate something, the channel determines performance. Fifth, this loss couldn't be explained by test-set contamination alone. Sixth, light fine-tuning didn't erase it either. Seventh, giving the model a larger thinking budget almost fully recovered the loss from typing errors, but did nothing for the voice-side loss — and for compressed speech, it actually got worse.
What This Means in Practice
If you're a solo operator juggling multiple roles, three changes are worth making. First, for tasks that require generating something new — drafting a proposal, reviewing terms and conditions, writing code — it's safer to give instructions by typing whenever possible. A few typos are fine; the model absorbs them on its own. On the other hand, for classification-style questions where the AI just picks among options, voice input on the go works perfectly well.
Second, when you do give instructions by voice, it's worth checking how much your dictation tool polishes your speech. A feature that smooths your words into tidy sentences can actually erase the tokens that carried your original intent, chipping away at voice-input accuracy. Leaving the transcript rough — closer to exactly what was said — may actually be safer for the AI. Third, keep in mind that the habit of switching on reasoning mode when a result looks off only helps with typed input; distortion introduced through speech wasn't recovered that way.
Caveats
This study used variations simulated with HIVE rather than recordings of real users' speech. Based on the abstract, details such as the number of models tested or the exact magnitude of the accuracy drop aren't disclosed, and Korean-language transcription quality may be a separate variable worth considering. Even so, the direction is clear: the channel through which you give AI its instructions is itself a variable in output quality, and there are tasks where writing has a real edge over speaking. If a task you handed off by voice today came back underwhelming, it's worth typing the same instruction back in for comparison.
