We Analysed 6,274 Voice Answers. 71% Were People Reading Scripts Aloud

Most voice answers on our platform are not opinions. Of the 6,274 voice answers people recorded on Would You Rather between 18 April and 8 October 2026, 4,467, or 71%, were people reading a script aloud for a speech dataset. Only 1,807 were spontaneous answers to open questions. Any "voice vs typed answers" claim built on the raw total would mostly be measuring how long it takes to read a sentence.

We went looking for a length multiplier and found a composition problem instead. This post shows what our voice answers actually are, what the same respondents typed elsewhere on the same form, and why we decided not to publish a "spoken answers are X times longer" number of our own.

What we counted and what we left out

Every number in this post comes from read-only queries we ran against Would You Rather's production database on 8 October 2026. The window runs from the first stored answer on 18 April 2026 to that date. We publish aggregates only: no form titles, no respondent details and no quotes from answers.

The raw database held 37,940 answer rows across 1,660 responses and 146 forms. Many rows are empty by design, because the form saves a row for section headers and for questions a respondent never reached. After removing empty rows and 4 test forms (single-response forms titled as tests or left untitled), we had 25,901 real answers. 6,274 of them carried a recording.

Three kinds of form, not one

We sorted every form with at least one response into three groups by what its questions ask for. The split matters more than any average in this post.

One more fact shapes everything below. 1,656 of the 1,660 responses belong to forms owned by a single account, which runs recruiting and speech-data projects in Indian languages. This is a study of how one heavy user's respondents behave, not a survey of our customer base.

71% of our voice answers were script reading

Script-reading forms are a different animal from surveys. A respondent sees a sentence, presses record, reads it, and moves on. A single person may record dozens or hundreds of sentences in one sitting. The median script-reading form in our data had 8 voice questions. The largest had more than 1,000.

That is why 116 responses produced 4,467 recordings, while 1,157 responses on open-question forms produced only 1,807. Per respondent, script reading generated about 25 times as many recordings. Any aggregate that mixes the two is dominated by the readers.

Most script recordings were in Indian languages

The reading forms were overwhelmingly multilingual. By the language each question was tagged with, or the script its sentence was written in, about 1,711 recordings were Hindi sentences in Devanagari and 1,032 were Malayalam. Tamil added 411, Odia 120 and Assamese 113, with a few dozen in Telugu and Kannada. English accounted for 473, and 573 more came from untagged sentences written in the Latin alphabet.

That mix tells you who reads aloud into a phone form today: people building speech datasets for languages that commercial recognisers still handle less well. We think that is a more honest description of where voice forms earn money right now than the customer-feedback use case most marketing leads with. If that is your use case, our guide to building a speech dataset for ASR covers the collection side.

Why script reading breaks the comparison

A reading task fixes the content. The respondent cannot say more or less than the sentence on screen. Our sampled script recordings had a median length of 10.3 seconds, which tells you about sentence length, not about voice as an answer format. We cover those recordings separately in how long people talk when a form asks a question.

We think most published voice-answer statistics from vendors suffer from the same blend. If a tool's customers include speech-data buyers, the headline count of "voice responses" is mostly read speech. Ask what the recordings were before you trust the average.

The same respondents, typing and speaking

The open-question group is where the interesting comparison lives. Almost all of it comes from one application form that 918 people completed. It asked a mix of tap questions, short typed answers and two spoken questions. Because the same people answered all of them, it is the closest thing we have to a within-person view.

The spoken questions were demanding. The first asked people to explain a technical concept in English, in 3 to 5 sentences, with a 12-second minimum and a 2-minute cap. The second asked them to explain a step-by-step everyday process in their own regional language, with a cap above 8 minutes.

What they typed

The typed questions on the same form were short-answer by design. Asked for their college and graduation year, people typed a median of 4 words. Asked for their current profession, they typed a median of 2. An optional long-text task asked them to listen to a clip and transcribe it, and those answers ran to a median of 89 words.

What they said

We sampled 140 recordings from each spoken question and measured them by decoding the audio. For the English explanation, the median was 38.2 seconds, with half of answers between 28.9 and 53.3 seconds. For the own-language explanation, the median was 61.7 seconds, with half between 40.2 and 111.8 seconds.

At conversational speeds, 38 seconds of speech is far more than 4 words. We still won't turn it into a word count. We don't know each respondent's speaking rate, and recordings include pauses and false starts. A precise-looking multiplier would be invented, not measured.

Why we won't publish an "X times longer" number

A fair length comparison needs three things we don't have. It needs the same question asked both ways, people randomly assigned to typing or speaking, and the same unit on both sides. Our form asked different questions in each format, nobody was randomised, and we have seconds for voice and words for text.

The last gap is a product fact. Would You Rather stores recordings, plays them in results and links them in exports, but it does not transcribe them. Without transcripts, we can't count spoken words. We would rather say that than estimate a number that ends up quoted without its caveats.

The controlled evidence still points one way

The best study we know is a 2024 experiment by Jan Karem Höhne, Konstantin Gavras and Joshua Claassen, published in Social Science Computer Review. They randomly assigned 1,001 German smartphone respondents to type or speak answers to four open questions. Spoken answers averaged 54.7 words against 21.5 typedHöhne, Gavras and Claassen 2024, Social Science Computer Review. N = 1,001 smartphone respondents, Germany. Mean words per given answer across four open questions., and covered more distinct topics.

Speed explains part of it. A lab study by Ruan and colleagues found that speech input reached 153 words per minute in English against 52 for a phone keyboardRuan, Wobbrock, Liou, Ng and Landay, arXiv 1608.07323, 2016. Lab text-entry study on an iPhone. English: 153 WPM speech vs 52 WPM keyboard.. That measured transcription of set phrases, not open answers, so treat it as direction rather than size.

What the voice answers tell us anyway

Even without words, the recordings carry signal. Three patterns held across our sample, and each one is useful when you design a form.

People talk longer in their own language

The own-language answers ran a median of 61.7 seconds against 38.2 for English. That gap is real in our sample, but it is confounded. The two questions asked different things and had different time caps, so we can't attribute the difference to language alone. We think it still argues for letting people answer in the language they think in, which we cover in collecting answers in Hindi, Tamil or any language.

Self-rated fluency didn't change answer length

The form asked people to rate their own English from 1 to 10. Answer length barely moved with it. Median English answers ran 38.9 seconds for people rating themselves 6 or below, 37.9 for 7 to 8, and 38.6 for 9 to 10Would You Rather, 137 sampled English recordings with a self-rating. Median seconds: rating 1 to 6, 38.9 (n = 14); 7 to 8, 37.9 (n = 43); 9 to 10, 38.6 (n = 80).. The low-rating group had only 14 recordings, so this is a hint, not a finding.

Only a few people hit the cap

On the English question, 9 of 138 sampled recordings ended within 2 seconds of the 2-minute limit. Most people stopped well short of it, which suggests the cap was generous rather than binding. That matters for anyone worried that a time limit cuts answers off.

Where this analysis is weakest

The biggest weakness is concentration. Almost every response came from forms run by one account, and the open-question figures rest on essentially one application form. Applicants trying to get work behave differently from customers giving feedback. They have a strong reason to finish and to speak at length.

Second, the duration figures come from a random sample of 140 recordings per question, not every file. Four of the 400 sampled files could not be decoded and were dropped. Third, recording length includes silence, so a long pause inflates the number.

There is also a gap we caused ourselves. Before 1 July 2026, 74 of 636 completed applications reached the database without at least one of their two required recordings. That was most likely a product fault rather than a choice: until then, recordings uploaded only at final submit, and the server did not re-check required questions. Since we started uploading each recording as it is made and blocking the Next button until it saves, none of 282 completed applications has arrived with a missing recording. The early gap means our sample slightly under-represents April to June respondents. We cover the full story in which gets skipped more, voice or text.

Fourth, and most important, nothing here is an experiment. The typed and spoken questions asked different things. If you want to know whether voice beats text for your audience, you need to randomise, and we explain the simplest version in the last section.

How to run your own voice vs typed test

You can get a better answer than ours with one form and two links. Keep it small, keep it honest, and decide what you will measure before you look at the results.

Before you start, read why the 'any other comments' box stays empty and why open-ended answers suffer when people must type. Both explain which questions deserve a microphone. For the drop-off side of the trade, see where respondents quit a form by question type.

We would not run this test on a cold audience with voice as the only option. Offer the typed box alongside the microphone in production, and keep the voice-only arm for the experiment. Our broader view of the trade-off is in our guide to voice surveys.

References

1. Höhne, J. K., Gavras, K. and Claassen, J., Typing or Speaking? Comparing Text and Voice Answers to Open Questions on Sensitive Topics in Smartphone Surveys, Social Science Computer Review, 2024. Used for words per answer and break-off by answer format.
2. Ruan, S., Wobbrock, J. O., Liou, K., Ng, A. and Landay, J. A., Speech Is 3x Faster than Typing for English and Mandarin Text Entry on Mobile Devices, arXiv, 2016. Used for speech vs keyboard entry speed.
3. Would You Rather, production database, read-only queries and a 400-recording audio sample, 8 October 2026. Used for every first-party figure. Aggregated and anonymised; not independently audited.

The weakest point in this source base: our first-party numbers come overwhelmingly from one account's forms, and the controlled comparison rests on one German experiment with a paid online panel.