Methodology
Voice Capture tells you that spoken survey answers carry more words than typed ones. We have measured that twice, and the two measurements did not agree: pilot testing before the product existed put it at 6×, and a later count over real client fieldwork put it at about 2×. Both are on this page, with the arithmetic behind each, and — the part that matters most — what neither of them can tell you.
The "6× more words" figure on the rest of this site comes from pilot testing that predates Voice Capture. Before there was a product to measure, we compared spoken against typed answers to the same questions; spoken answers came back around 6× longer, and that is what convinced us this was worth building.
One pair from that testing, both answering the same question: the spoken answer ran 47 words, the typed one 6. It is a single illustrative pair, not an average — we are showing you what the difference looked like, not summarising the pilot with it.
What this page cannot show you about the pilot: how many responses it covered, and exactly when it ran. That testing came out of a client engagement, so its size and timing are not ours to publish — this is a confidentiality boundary, not a detail we forgot to write down, and it is not going to be filled in later. It has a cost worth naming: the headline cites a source you cannot inspect. That is why the production measurement below is published in full, sample and all. It is the number you can actually hold us to.
Once the product was live we ran the same comparison properly, over real client fieldwork, at a scale the pilot never had. It came out lower: about 2×, not 6×. We publish it anyway, in full, because a number that got smaller when it got more rigorous is worth more to you than one that only ever got quoted.
The two are not measuring quite the same thing. The pilot was small, informal and self-selected — our own questions, chosen because they were the kind of open-end voice ought to help most. What follows is every eligible open-end across seven client studies, good questions and dull ones alike, with no say in which got asked. A large, unselected count regressing toward a smaller number than a small, hand-picked one is the ordinary shape of this, not an anomaly — and it is why the figure below is the one with the receipts.
Words per response, voice against typed, in real client surveys. In every study included here both answer modes were live on the same questionnaire: some respondents pressed the microphone button, others typed into the box. We counted the words in each answer and compared the two groups. Nothing is modelled, weighted or extrapolated — these are counts.
693 voice responses and 6,476 typed responses, 7,169 in total, across 7 client studies fielded between March 2026 and August 2026.
Voice answers average 23.1 words against 10.1 typed, a ratio of 2.3×. The median voice answer is 16 words against 8 typed.
We report medians alongside means on purpose. Both groups have long tails — a small number of very long answers pulls the average up — so the mean on its own would overstate the typical case. The median is the answer in the middle, and it is the more honest number for what a respondent usually says or writes.
| Metric | Voice | Typed |
|---|---|---|
| Responses (n) | 693 | 6,476 |
| Mean words | 23.1 | 10.1 |
| Median words | 16 | 8 |
| 90th percentile | 48 | 19 |
Both published figures are rounded down from the measurement they come from, so neither can ever be larger than its own evidence: the pilot is published as 6×, and this count as 2×. The rest of the site quotes the first; this page is where the second one lives.
The comparison above pools every question together, so we also ran it question by question. Across 9 question cells with at least ten responses in each group — 606 voice and 5,463 typed responses — the median ratio is 2.0× and the mean ratio 2.06×, ranging from 1.42× to 2.60×.
This is the check that matters: comparing voice against typed within the same question rules out the ratio being an artifact of different studies asking different questions, rather than of how the answer was given.
Voice Capture cannot measure skip or abandonment rates. It only records answers that happened: there is no record of a respondent who saw a question and did not answer it, so the denominator a skip rate needs does not exist in our data at all. That denominator lives in your survey platform.
This site previously claimed "40% fewer skips" and attributed it to our own data. No such measurement ever existed. The claim has been removed rather than restated with a smaller number.
We also make no claim here about the quality of what was said — only about how much of it there is. More words is more material to code and analyse; it is not automatically a better answer.
And this describes our clients’ fieldwork, not a controlled experiment. Respondents chose which mode to use; nobody was assigned to one.
Two kinds of responses are left out, under a rule written down before the figures were read:
Both rules, and the list of included studies, live in the source file these figures come from, so a re-measurement reproduces the same population instead of a new one.
This site once published a words-per-answer figure with no source anywhere in our code, our data or our documents. It survived for months because nothing tied it to a measurement. Every number on this page is now rendered from the single module that holds both measurements, and so is every number in the site copy that cites them. A figure cannot be edited inside a sentence without changing the evidence, and it cannot change in the evidence without changing the sentence — which is also why we could not simply make the two numbers agree, and did not.
Add the microphone to one open-ended question in your next study and compare the two columns yourself.
Start free