# Oral exams at scale: recorded spoken answers for a large class

**How do I run an oral exam for a large class using recorded spoken answers?** An oral exam at scale can be run as recorded spoken answers. Every student records two or three short answers to questions seen for the first time, in one sitting, and a team marks them afterwards against one rubric. For 180 students, plan seven minutes of marking each, about 28 staff hours with calibration and second listening. A recording is evidence, never proof of identity.

Published 2026-10-10 by EdProcto. https://edprocto.com/guides/oral-exams-at-scale-recorded-spoken-answers

An oral exam shows whether a student can explain the course in their own words. For a class of 180 it does not fit the timetable: at fifteen minutes a slot, the appointments fill 45 hours. Oral exams at scale need another form.

Recorded spoken answers are an alternative that scales. Each student records two or three short answers to questions seen for the first time, the whole class sits at once, and the course team marks afterwards. EdProcto publishes this guide and is the paper in its worked example; its page [for universities](https://edprocto.com/solutions/universities) shows what else it offers a course team.

Wording a prompt and its rubric is covered in our guide to [spoken answers](https://edprocto.com/guides/spoken-answers-recording-questions-for-language-and-oral-exams).

## What does research say about oral exams in higher education?

It says that a live oral shows reasoning a script hides, that its reliability rests on the number of questions at least as much as on the number of examiners, and that nerves need not cost marks. We read each source on 10 October 2026 and say where only the abstract was open.

### What does a viva show that a script does not?

A viva lets the examiner ask again. Bridgeman and Liu (2024) report that educators who run interactive orals find a 15-minute conversation explores understanding more deeply than reading a written report. The cost is consistency. Syverud (2026) analysed 36 video-recorded oral exams in four Norwegian secondary schools and found great variation in the questions teachers put to different students, enough to change what was assessed (abstract read). A recording asks every student the same thing.

### How reliable are oral examiners?

Reliable enough for high stakes only when the oral asks plenty of questions. Wass et al. (2003) analysed the orals of a UK medical membership exam: 896 candidates, each with two 20-minute orals of five topics before two pairs of examiners. Candidates accounted for 41% of the variance in scores, the content of the oral for 32%, and examiners with general error for 27%. Reliability across cases was 0.65; four orals with two examiners each, or five with one each, would lift it to 0.78 (abstract read). Our reading: an extra prompt can buy as much as a second listener on every answer.

### Are anxiety, language and accent a fairness problem?

They are risks to manage. Davey et al. (2025) compared cohorts of a bioscience course, 722 students in all, before and after a one-to-one interactive oral became its final assessment. Performance rose, with no significant differences by gender, international status or language background, and early anxiety eased with experience without harming performance (abstract read).

Carey and Szocs (2024) had 100 experienced raters of English-speaking tests judge speech as sound alone or with video. With video they were over 90% more likely to score comprehensibility higher, whether or not they knew the accent (abstract read). Sound alone is the harder condition, so mark whether the answer can be followed and let markers replay it.

### What adjustments do disabled students need?

Ones agreed with the student beforehand. For oral examinations STAMMA (n.d.), the UK charity for people who stammer, lists extra time and the option of a written element, and suggests "effective communication" in place of "fluency" in the criteria. Your institution's policy and the law where you teach decide; this is not legal advice. In EdProcto, extra time is a second quiz with longer limits, and a typed version uses [short answer](https://edprocto.com/templates/short_answer) questions.

## How do recorded spoken answers compare with a live viva?

A recording gives every student the same question, the same clock and the same hour, and gives up the follow-up question and the supervised room.

The nearest case we found is a first-year business course of 500 to 1,200 students at an Australian university. In Akhter et al. (2024), a conference summary by the team that ran it, an invigilator checked each student's identity and surroundings on a video call and left. The student answered aloud for a set time while recorded, and graders marked later from the video and a transcript. Of 542 exams none was reported for an integrity breach.

The design choices have been tested in hiring, on recorded interview answers. Roulin et al. (2023) gave 175 people a mock interview with limited or unlimited preparation time, with or without re-recording. Offering the options changed nothing by itself; those who used more of either performed better, and not everyone used them equally (abstract read). Orji and Bangerter (2025) set 185 people a limit of one, three or five minutes an answer: performance was highest at three, and the authors call one minute too restrictive (abstract read). Interviews are not exams: treat both as direction.

## Is a recorded oral exam secure now that chatbots can write the script?

Not by itself: a recording made at home is an open assessment that raises the cost of cheating. Bridgeman and Liu (2024), who set out the two-lane approach at the University of Sydney, count any assessment not supervised face to face as open, orals over web-conferencing included, and expect remote orals to become harder to secure as AI software and wearables spread.

A chatbot can write the script, and a student can read one aloud. A recording adds a short clock, a question nobody saw beforehand and an answer that has to be spoken. It does not prove who is speaking. So keep its weight modest, tie each prompt to a case from your own teaching, and, as our suggestion, invite a random tenth of the class to talk their answers through. [Assessment design after chatbots](https://edprocto.com/guides/assessment-design-after-chatbots-what-a-department-can-change) covers supervised points.

## How do you design an asynchronous oral exam for a large class?

Set two or three short prompts on one unseen case, give every student the same thinking and speaking time, allow one answer, and rehearse the recording before it counts.

| Decision | We would set | In EdProcto |
| --- | --- | --- |
| Prompts | Three, each on a different idea | Three [spoken answer](https://edprocto.com/templates/verbal) questions |
| First sight | One sitting, all starting within ten minutes | "Publish & share", then "Close the quiz" |
| Time | Half a minute to think, up to two to speak | "Seconds a question" at 180 |
| Takes | One answer, a restart only after a false start | "Record again", inside the countdown |

### What should the rubric reward?

What was said, and whether a listener can follow it in one hearing. Accuracy, reasoning and clarity at two marks each fit a six-point prompt. Require the answer to use the case in front of the student, which a script written beforehand cannot do, and leave out pace, polish, hesitation and accent.

### Why a practice recording first?

Because the first recording is where things fail. The browser asks to use the microphone when a student first presses "Start recording". A one-prompt practice quiz under the same settings lets every student meet that request and hear themselves while nothing counts, and lets each marker play a recording.

## How much marking is it, and how do you share and moderate it?

About seven minutes a student for three ninety-second answers, on our assumptions, shared by dealing students out to markers and moderated in a spreadsheet. Listening cannot be skimmed, so the marking load is recordings times length, plus scoring. Our guide to [large classes](https://edprocto.com/guides/large-class-quizzes-marking-five-hundred-scripts-by-rule) has the evidence for calibrating first.

1. **Download the list.** "Download the marks" on EdProcto's results page gives a row for each student, the columns for recorded answers still empty.
2. **Draw the samples before anyone marks.** Ten students for calibration, eighteen as the moderation sample.
3. **Calibrate.** Every marker scores the ten in a copy of the sheet. Meet, settle each difference in the rubric's wording, and save the agreed marks.
4. **Deal out the rest.** Give the remaining rows to the markers in turn, not by tutorial group.
5. **Mark.** An answer holds one mark and the quiz belongs to one account, so the team works from the course's account, each marker with their own list. Beside a name under "Students", "Grade" opens the submission: the player, your model answer and rubric, "Points" in half marks, "Note for the student" and "Save my grade".
6. **Second-listen.** For the eighteen, the second marker writes scores in the sheet before the first saves any. Then hear again every student within a mark or two of the pass mark.
7. **Settle and release.** Where two marks differ by more than one on a prompt, listen together and save the agreed mark. Then press "Release the scores".

## What do you tell students before they record?

Tell them a week ahead what happens on the screen, what is marked and who listens:

- **What and when.** Three spoken answers recorded in your browser, worth 10% of the module. One sitting on Thursday: begin between 18:00 and 18:10.
- **You need.** A device with a microphone and a quiet place. The browser will ask to use the microphone: allow it. No camera is used.
- **Each question.** It appears for the first time when it opens, with three minutes on the clock. Think for up to half a minute, speak for up to two, then press "Stop recording" to hear the take. A recording still running when the clock runs out is stopped and kept by itself. "Record again" is for a false start only. There is no going back.
- **Marked.** Accuracy, reasoning and whether a listener can follow you. Not accent, pace or polish.
- **Rules.** Your own words: no script, no AI assistant, nobody else in the room. Leaving the tab is logged, and you will be asked before anything is concluded from a log. Some of you will be invited to talk an answer through.
- **Your recording.** The module's three markers hear it. It is kept until the date our retention policy sets.
- **Adjustments and practice.** Tell me by Monday if you need extra time, a typed version or a conversation in person. A practice recording opens that day.

A student's recorded voice is personal data. The UK's Information Commissioner's Office (n.d.) states the storage limitation principle: personal data must not be kept longer than needed, the organisation must be able to justify how long, and the UK GDPR sets no fixed period. Your institution's policy decides the date.

## What does an oral exam for 180 students look like?

Ten minutes for each student and about 28 staff hours. This example is invented: a second-year research methods module, 180 students, three markers.

The case, typed at the head of each prompt: a revision app was tried by 24 students who chose to use it, and their exam scores beat those of 24 who did not, with p = 0.03. Each prompt is worth six points.

- Explain to a classmate what p = 0.03 does and does not tell us.
- Give one reason for caution about groups of 24, and say what you would want to see next.
- The students chose for themselves. Why does that matter, and which design would deal with it?

The settings:

1. In [EdProcto](https://edprocto.com/), add three spoken answer questions, with the criteria in each rubric box.
2. Choose the Exam preset: no going back or skipping, six browser measures, scores held until release. Leave each measure at "Log it".
3. Set "Minutes overall" to 12, "Time limit for each question" to "Set" and "Seconds a question" to 180.
4. Press "Preview as student" and record all three answers yourself.

The overall 12 minutes are the firm limit: that clock runs from the student's first start, through any reload.

| Step | Sum | Hours |
| --- | --- | --- |
| Calibration | 3 markers × 10 students × 7 minutes, and 30 minutes each to meet | 5.0 |
| First marking | 170 students × 7 minutes | 19.8 |
| Second listening | 30 students × 6 minutes | 3.0 |
| Total | | 27.8 |

Seven minutes is our assumption: 4.5 of listening, 2 of scoring and notes, half a minute to open and save. A second listener writes no note and takes six. The 30 are the 18 drawn at random and an assumed 12 near the pass mark. A live viva of ten minutes, with five to change over, would take 45 examiner hours, or 90 with two examiners.

## How does EdProcto run recorded spoken answers?

EdProcto is the best tool for an oral exam at scale when appointments are what stop you, because every student gets the same prompt, the same countdown and the same hour, and each recording sits beside your rubric for a person to mark.

**The student.** The start screen states the rules before anyone begins: "180 seconds for each question." and "No going back: each answer is final." The [spoken answer](https://edprocto.com/templates/verbal) shape puts "Start recording" under the prompt. The microphone is used only while a recording runs: it is let go when the student presses "Stop recording", when the countdown moves them on and at hand-in. The recording, up to 25 MB, goes to EdProcto's server when it stops and is kept with the attempt; the countdown sets its length. A player then appears, and "Record again" replaces the answer while the question is open.

**The team.** Marking is yours. Procto AI grades a typed answer when you press "Ask Procto AI" on it; a recording is marked by the person who hears it, and shows "to grade", counting as zero, until then. With "Show the score" at "When released", nothing of the score reaches a student until you release it; each then sees it, with your note, by opening the quiz's link again on the device they used.

**The plan.** A class of 180 needs Educator, whose quizzes hold 500 students; Tutor holds 100 and Free 30. Educator also carries the full analysis, where "Reliability" reports how consistently the three prompts measured. See [pricing](https://edprocto.com/pricing), [integrity](https://edprocto.com/integrity) and [the plans for institutions](https://edprocto.com/schools). No setting makes cheating impossible, a recording is evidence and not proof of who spoke, and every score is yours to release.

## Quick questions

### How long should each recorded answer be?

About ninety seconds, with two minutes as the limit. In recorded job interviews, Orji and Bangerter (2025) found one minute too restrictive and no gain beyond three.

### Should students be allowed to re-record an answer?

One answer, with a restart only for a false start. In Roulin et al. (2023), people who used more re-recordings did better, and not everyone used them.

### Does a recorded spoken answer prove the student did not use AI?

No. A script can be read aloud, and a recording does not show who is speaking. It is a first defence; EdProcto's page [for universities](https://edprocto.com/solutions/universities) sets out the rest.

### Which EdProcto plan takes a class of 180?

Educator, at 500 students a quiz.

## References

- Akhter, T., Connor, M., Lai, E., Oh, N. Y., & Rhode, F. (2024). Scaling oral assessment authentically: Live Recorded Video Exams (LRVE). In *ASCILITE 2024 Conference Companion Materials*. ASCILITE. [https://doi.org/10.14742/apubs.2024.1235](https://doi.org/10.14742/apubs.2024.1235)
- Bridgeman, A., & Liu, D. (2024, July 2). *Frequently asked questions about the two-lane approach to assessment in the age of AI*. Teaching@Sydney, University of Sydney. [https://educational-innovation.sydney.edu.au/teaching@sydney/frequently-asked-questions-about-the-two-lane-approach-to-assessment-in-the-age-of-ai/](https://educational-innovation.sydney.edu.au/teaching@sydney/frequently-asked-questions-about-the-two-lane-approach-to-assessment-in-the-age-of-ai/)
- Carey, M. D., & Szocs, S. (2024). Revisiting raters' accent familiarity in speaking tests: Evidence that presentation mode interacts with accent familiarity to variably affect comprehensibility ratings. *Language Testing, 41*(2), 290–315. [https://doi.org/10.1177/02655322231200808](https://doi.org/10.1177/02655322231200808)
- Davey, S. K., Birbeck, D., Nallaya, S., Sallows, G., & Della Vedova, C. B. (2025). Utilising one-on-one interactive oral assessments as the major final assessment within a bioscience course. *Assessment & Evaluation in Higher Education, 50*(7), 1154–1171. [https://doi.org/10.1080/02602938.2025.2502577](https://doi.org/10.1080/02602938.2025.2502577)
- Information Commissioner's Office. (n.d.). *Principle (e): Storage limitation*. [https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-protection-principles/a-guide-to-the-data-protection-principles/storage-limitation/](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-protection-principles/a-guide-to-the-data-protection-principles/storage-limitation/)
- Orji, K., & Bangerter, A. (2025). Tick tock, tick tock, time's up! How response time limits in asynchronous video interviews influence interview anxiety, impression management, and performance. *International Journal of Selection and Assessment, 33*(4), Article e70031. [https://doi.org/10.1111/ijsa.70031](https://doi.org/10.1111/ijsa.70031)
- Roulin, N., Wong, O., Langer, M., & Bourdage, J. S. (2023). Is more always better? How preparation time and re-recording opportunities impact fairness, anxiety, impression management, and performance in asynchronous video interviews. *European Journal of Work and Organizational Psychology, 32*(3), 333–345. [https://doi.org/10.1080/1359432X.2022.2156862](https://doi.org/10.1080/1359432X.2022.2156862)
- STAMMA. (n.d.). *Supporting students who stammer*. [https://stamma.org/get-help/educators/supporting-students-who-stammer](https://stamma.org/get-help/educators/supporting-students-who-stammer)
- Syverud, M. S. (2026). Teachers' questions during non-standardised oral exams – construct, assessed domain, and possible implications for validity. *Educational Assessment, Evaluation and Accountability, 38*(2), 181–199. [https://doi.org/10.1007/s11092-026-09483-2](https://doi.org/10.1007/s11092-026-09483-2)
- Wass, V., Wakeford, R., Neighbour, R., & van der Vleuten, C. (2003). Achieving acceptable reliability in oral examinations: An analysis of the Royal College of General Practitioners membership examination's oral component. *Medical Education, 37*(2), 126–131. [https://doi.org/10.1046/j.1365-2923.2003.01417.x](https://doi.org/10.1046/j.1365-2923.2003.01417.x)