What Is a Voice Survey?

A voice survey is a survey in which respondents can answer open-ended questions by speaking instead of typing. The audio is transcribed automatically, and the text arrives in your data file alongside every other answer.

That definition is deliberately narrow, because "voice survey" has become a label for three quite different things. Choosing between them is the first real decision in a voice research project, and they are not competing versions of one product — they answer different questions, and they cost very different amounts of your time.

Voice inside a survey you already run

A record button sits next to the open-ended questions of an ordinary online survey. Everything else stays as it is: the questionnaire, the routing, the quotas, the panel, the export. The respondent speaks, and the transcript lands in the same field a typed answer would have gone into. This is the lowest-friction option and the only one that drops into a tracker or an omnibus wave without renegotiating anything.

AI-moderated interviews

A separate platform runs a conversation: it asks, listens, and decides what to ask next, following up where an answer is thin or unexpectedly interesting. For exploratory work that would otherwise need a human moderator, this is genuinely powerful. The trade is that the study lives on that platform — you rebuild the guide there, recruit into it, and reconcile its output with the rest of your data afterwards. If your research question is "what don't we know yet", this is often the right call, and it is worth saying plainly that putting a microphone on a fixed questionnaire is not a substitute for it.

Phone and IVR surveys

The oldest form: a call, with either a live interviewer or an automated script. Still the right tool for populations that are not reliably online, and still the most expensive way to buy a completed interview. Worth separating from two adjacent terms that sound identical: "voice of the customer" usually describes a feedback programme rather than a data collection method, and "voice analytics" usually means mining recordings of call-centre conversations that were going to happen anyway, not fielding a study.

The rest of this guide is about the first kind, because it is the one most market research teams can adopt this month without changing anything else about how they work.

Why Voice Data Is Different for Open-Ends

The core advantage is volume, and it is worth being precise about how much. We measured it on our own production data: 693 voice and 6,476 typed open-end responses across seven client studies fielded between March and August 2026. Spoken answers averaged 23.1 words against 10.1 typed, with medians of 16 and 8. Compared within the same question, the ratio holds at about 2.0 times. Roughly twice as long, then — a real difference, but not the order-of-magnitude gap this category likes to advertise (full methodology and sample).

The more interesting question is what the extra words contain, and here the honest answer is that length is what we measured and richness is what we observed. Read a few hundred transcripts next to their typed equivalents and the same patterns recur:

Reasons, not just verdicts. A typed answer says "too expensive". A spoken one says it is too expensive compared to what, for whom, and at what moment the person decided.
Hedging that carries meaning. "I mean, it's fine, but..." is a different data point from "It's fine", and typing tends to sand that off.
Unprompted associations. People talk their way into connections they would have edited out of a written box, because writing is composed and speech is discovered.
Ordinary language. Respondents describe products the way customers actually describe them, rather than in the clipped register people adopt for survey text fields.

One caveat we would rather state ourselves: more words is more material to code and analyse. It is not automatically a better answer, and we make no claim about the quality of what is said — only about how much of it there is. A rambling ninety seconds can carry less insight than one precise typed sentence. What voice reliably changes is the floor, not the ceiling.

When to Use Voice Surveys

The verbatim is the deliverable

If your report will feature customer quotes, voice gives you quotes that sound like a person talking. Typed open-ends produce material that is accurate but flat, and a deck full of four-word verbatims is a hard sell to a client who wanted to hear their customer.

Your respondents are on phones

Typing a considered paragraph on a phone keyboard is work, and speaking is not. We should be careful here: we have not measured completion or drop-out effects, and our own data cannot measure them, because it contains only responses that happened. What we can say is that the effort asked of a mobile respondent is lower, and that the answers we do receive are longer.

The topic carries emotional weight

Health, bereavement, money trouble, workplace conflict. When a subject is difficult, composing sentences in a text box adds a second task on top of the hard one. Speaking is closer to how people process those topics anyway.

You are testing creative or advertising

Post-exposure open-ends after a video ad are among the best uses of voice. You want the immediate reaction, not a written paragraph composed after the respondent has had three more screens to rationalise it.

Your respondents are time-poor and hard to reach

Clinicians, executives, specialists on a panel you pay a premium for. Anything that lowers the cost of answering well is worth more with this group than with a general population sample, simply because each completed interview is worth more.

When Text Is Still Better

Voice is not always right, and a guide that pretends otherwise is selling something. Stay with text when:

The answer is structured. Codes, numbers, product names, addresses, anything where exact spelling matters. Transcription of a rattled-off model number is a bad bet.
The respondent is in public. Commuters, open-plan offices, anyone answering a survey somewhere they would rather not be overheard.
The subject is socially sensitive in a way speech makes worse. Emotional weight often favours voice; social desirability often does not. Saying an unflattering thing out loud is harder than typing it, even alone in a room.
Your sample skews to segments that dislike speaking to a device. This varies enormously by market and age, and it is worth a soft-launch check rather than an assumption.

In practice the answer is rarely voice or text. It is voice offered alongside the text box, with respondents self-selecting. You get longer answers from the people who prefer speaking and lose nothing from the people who do not.

How to Choose Voice Survey Software

Most comparisons of voice survey tools compare features. The features converge quickly; the architecture does not, and the architecture is what determines how much work the study becomes. Five questions separate the options faster than any feature grid:

1. Does it live inside your survey, or replace it?

This is the question that decides the others. A tool that embeds into your existing platform inherits your questionnaire, logic, quotas, panel and data file. A tool that hosts the study means rebuilding all of that in a second system and reconciling two datasets afterwards. Neither is wrong — but the second is a project, and the first is an afternoon.

2. Where does the transcript end up?

The good answer is "in the response record, in the same row as every other answer, in your normal export". The expensive answer is "in a separate dashboard you export from and join by respondent ID later". Joining sounds trivial until you are doing it every wave of a tracker.

3. What happens to the audio?

Ask whether recordings are stored, where, for how long, and whether the vendor will sign a data processing agreement. This is the question your legal or compliance team will ask you eventually, and it is much easier to answer before fieldwork than after. Voice Capture processes audio in memory and retains only the transcript, on EU-hosted infrastructure, with a published DPA.

4. What does it do in the languages you actually field in?

Automatic speech recognition is not uniformly good across languages, accents and recording conditions. If you run multi-market work, test your hardest market first rather than your easiest — the pilot that reassures you least is the useful one.

5. How does it price, and what happens if a wave is small?

Per-response pricing suits studies with unpredictable volumes; seat or platform licensing suits continuous programmes. The mismatch to avoid is paying a subscription through the months a tracker is not in field. Our own pricing is one-time credit packs for this reason: credits do not expire, so a quiet quarter costs nothing.

Designing Voice Survey Questions

Voice questions are not text questions with a microphone bolted on. The goal is natural speech, and the way you ask determines whether you get a narrative or a recitation.

Write it the way you would say it

Instead of "Please describe your overall experience with the product." try "Tell us about your experience with the product, like you are talking to a friend." The first invites a formal report. The second invites a story. Read every voice prompt aloud before fielding it; anything that feels strange to say is going to feel strange to answer.

Give explicit permission to ramble

A short lead-in does real work: "There are no right or wrong answers, and it does not need to be tidy. Just talk for about a minute." Without it, respondents perform survey answers into the microphone and you get spoken versions of typed brevity.

Say how long you want, and cap it

Naming a duration is the single cheapest improvement to voice answer quality, because "about a minute" tells the respondent what shape of answer you want. Cap the recording at 90 to 120 seconds. A limit is not a constraint on richness; it forces prioritisation, and it keeps your analysis tractable.

Ask one thing per recording

A reader can scroll back up to a two-part question. A speaker cannot. Compound questions get half-answered, and it is always the second half that goes missing. If you have two things to learn, use two prompts.

Put the voice question where energy is highest

Early enough that the respondent has not fatigued, late enough that they have context. Immediately after the stimulus, or immediately after the rating whose reasons you want, is usually right. The worst placement is the last question of a long grid-heavy survey.

Never remove the text box

Some respondents will not or cannot speak, and forcing the modality costs you their answer entirely. Offering both also gives you a natural comparison within the same question, which is exactly how we measured the difference between the two.

Setting Up a Voice Survey

For most market research teams this is an existing survey plus two copy-and-paste steps: a script tag in the survey's custom HEAD, and a short action naming the question you want a record button on. No development, and nothing about the questionnaire changes.

The platform-specific walkthroughs are here: Alchemer (still called SurveyGizmo by plenty of teams), QuestionPro, or the generic embed for any web form that accepts custom HTML. There is also a longer step-by-step Alchemer guide with screenshots.

One piece of fieldwork advice that has nothing to do with the software: soft-launch. Send ten percent of the sample, read the transcripts yourself, and check the prompt is producing the kind of answer you expected before the rest goes out. Voice prompts fail differently from text questions, and they fail in ways you can hear immediately.

What to Expect in Field

Three things surprise teams running voice for the first time, and all three are manageable if you plan for them.

The browser asks for microphone permission. This is a one-time prompt and it is the main place respondents drop the interaction, so say what is coming before it appears: a line of survey text explaining that the next question can be answered by speaking, and that the browser will ask for access, removes most of the friction.

Not everyone will speak, and that is fine. Voice is an option, not a mode you assign. Expect a mix, expect it to differ by market and by device, and design your analysis so that a text answer and a voice answer to the same question sit in the same column and are comparable.

Noise is a fieldwork variable, not a bug. Real respondents answer in kitchens, on trains and in offices. Modern transcription copes with a great deal of this, but a very short answer that transcribes as gibberish is usually an environment problem rather than a software one, and it is worth flagging those cases in QA rather than treating them as data.

Analyzing Voice Survey Data

Analysis follows the same path as any open-end. The difference is that you start from an automatic transcript rather than something the respondent typed.

Step 1: Read a sample before you code anything

Transcription is very good and not perfect. Scan a sample for systematic errors, particularly proper nouns, brand names and category jargon, which tend to come back phonetically. Systematic is the key word: a brand consistently mis-transcribed the same way is a find-and-replace, while scattered single-word slips rarely change a code frame at all.

Step 2: Watch for confident nonsense on near-silent clips

Speech recognition models can produce fluent text from a recording that contains almost nothing — a respondent who tapped record and said nothing, a clip that is pure background noise. The output looks like a normal answer, which is what makes it dangerous in a code frame. Voice Capture flags these rather than passing them through silently, but whatever tool you use, check what it does with an empty recording before you trust a wave of data.

Step 3: Code with your existing framework

A transcript is an open-end. Your code frame, your coder instructions and your reliability checks all still apply. On larger datasets, AI-assisted coding tools such as Survey Coder can do the first pass on theme extraction, with a human reviewing the frame rather than building it from scratch.

One adjustment is worth making: a spoken answer that runs to twice the length of a typed one will often carry two or three codeable ideas rather than one. Frames written for terse typed open-ends tend to force a single code per response, and applied to voice data they quietly discard the second half of what the respondent said.

Step 4: Flag the quotes while you are in there

The practical advantage of voice for reporting is that transcripts read as speech. Tag the powerful ones as you code rather than hunting for them the week the deck is due.

Step 5: Merge back into the quantitative data

Because the transcript lives in the response record, it exports with everything else and joins on the respondent ID you already have. That is what makes voice themes cross-tabbable against your quant variables — theme by segment, by satisfaction score, by wave. Two places this pays off immediately: the follow-up to an NPS score, where "why did you give that number" is the entire point of the question, and ongoing customer feedback studies, where the same open end is asked wave after wave and short typed answers compound into a thin tracker.

Consent, Privacy and Retention

Voice responses are personal data, and a recording of someone's voice feels more personal to them than a line of typed text — a distinction worth respecting even where the law treats both the same.

Get the basics right and the rest is straightforward. State in the survey introduction that some questions can be answered by speaking, and say what happens to the recording. Keep voice optional with a text alternative always available. Specify your retention period in the same place you specify it for the rest of the study, and make sure your vendor's answer matches what you have told respondents.

For Voice Capture specifically: audio is processed in memory and never stored, only the transcript is retained, processing runs on EU-hosted infrastructure, and there is a published data processing agreement and a privacy policy you can hand to a compliance team without a phone call.

Where the Category Is Going

The interesting development in voice research is not better transcription — that problem is largely solved. It is that conversation itself is becoming automatable: platforms that listen to an answer and decide what to ask next, running something close to a moderated depth interview at panel scale.

Be clear about what that is and is not. Those are separate platforms, and the cost of the capability is that the study moves onto them. Voice Capture is narrower than that: it puts a microphone on the questions you already wrote, in the survey you already run, and an add-on can ask one AI follow-up on the answer to an open end you choose. What it does not do is moderate the session — it never decides what the study asks next. If you need an adaptive conversation from end to end, an AI-moderated platform is the honest recommendation, and we have written a comparison of the two approaches that says so at more length.

For the large amount of research that is a fixed questionnaire fielded to a panel with quotas — trackers, concept tests, brand equity waves, customer feedback programmes — the constraint was never the absence of a moderator. It was that the open end at the end of the grid came back with eight words in it.

Add Voice to Your Next Survey — Free →

Continue reading: Why Voice Data Captures What Text Surveys Miss | How to Analyze Voice Survey Responses with AI Transcription | How we measured the difference