Online Speaking Test: How to Assess Spoken Language Skills With a CEFR Rubric
IELTS Speaking takes 11 to 14 minutes with an examiner. How to run a low-stakes online speaking test with a CEFR rubric, and when you need a certified test.
The short version
- The Council of Europe's CEFR Companion Volume (2020) judges pronunciation by intelligibility, not accent, and says its top level has no relation to an idealised native speaker. Build your rubric the same way.
- Certified tests sample very little speech: IELTS Speaking is an 11 to 14 minute examiner interview, the TOEFL iBT Speaking section is about 8 minutes, and Pearson's Versant Speaking and Listening Test takes 17 minutes.
- A voice form can run a fair low-stakes screen: 3 or 4 spoken tasks, a 4-level CEFR-based rubric, and two independent listeners, at no per-candidate test fee.
- Use a certified test whenever the result decides a visa, an admission or a regulated job, or when identity checks and score comparability matter. A form has neither.
Accent is not a speaking skill. The Council of Europe rebuilt its pronunciation scale around that idea in 2020. It is the most useful thing to know before you design an online speaking test. Certified exams like IELTS spend 11 to 14 minutes listening to a candidate. You can run a fair, low-stakes version of that with a voice form, as long as you know where it stops being enough.
A good spoken English test online asks one question about pronunciation: could a listener understand without effort? Below is a rubric built from the official CEFR descriptors. Then a comparison of what IELTS, TOEFL and Versant actually test. Then a plain line between a screen and a certificate.
To assess spoken language skills online, give candidates 3 or 4 recorded speaking tasks, from a short personal answer to an opinion they must justify. Score each recording against CEFR descriptors for fluency, range, coherence and intelligibility, with two independent listeners. Use a certified test such as IELTS, TOEFL or Versant when the result has legal or admission weight.
What an online speaking test actually measures
A speaking test turns a few minutes of speech into a claim about real conversations. That is a big leap from a small sample. Every design decision, from task choice to rubric, exists to make the leap less risky.
Most serious speaking assessments look at the same handful of qualities. The CEFR Companion Volume, the Council of Europe's 2020 update to the framework, lists six for spoken language: range, accuracy, fluency, interaction, coherence and phonology. IELTS scores four related criteria. A rubric that drops any of these is measuring something narrower than speaking.
Proficiency is not the same as polish
A proficiency assessment asks what someone can do in the language now. It does not ask whether they sound like a newsreader. That distinction matters most in places like India, where many fluent professionals speak English shaped by one or more other languages.
The CEFR is direct about this. It says its top level, C2, has no relation whatsoever with the performance of an idealised native speaker. We think any employer rubric that quietly uses "sounds native" as its top band is measuring background, not skill.
Why the CEFR dropped accent as a marker of skill
The 2001 pronunciation scale jumped from accepting "a foreign accent" at B1 to expecting "clear, natural pronunciation" at B2. The Companion Volume replaced it. The authors wrote that focusing on accent and accuracy instead of intelligibility had been detrimental to teaching pronunciation, and rebuilt the scale from scratch.
In the new scale, intelligibility separates the levels. The key question is how much effort a listener needs to decode the message. Even at C1, the descriptor says features of accent from other languages may be noticeable but do not affect intelligibility. That is the standard we would hold any speaking screen to.
A CEFR speaking rubric you can use
You do not need all six levels for most screens. Hiring and course placement decisions usually turn on the line between A2 and B1, or between B1 and B2. The table below quotes or closely paraphrases the official descriptors for 4 levels and 4 qualities.
| Level | Fluency | Range | Coherence | Intelligibility |
|---|---|---|---|---|
| A2 | Understood in very short utterances; pauses, false starts and reformulation very evident | Basic patterns and memorised phrases for simple everyday situations | Links words with simple connectors like "and", "but", "because" | Listeners need to ask for repetition from time to time |
| B1 | Keeps going comprehensibly, with evident pausing to plan and repair | Enough language to get by on familiar topics, with some hesitation | Links shorter elements into a connected, linear sequence of points | Generally intelligible; accent usually influenced by other languages |
| B2 | Fairly even tempo; may hesitate, but few noticeably long pauses | Clear descriptions and viewpoints on most general topics | Uses a limited number of cohesive devices, with some "jumpiness" in long turns | Accent influenced by other languages, with little or no effect on intelligibility |
| C1 | Fluent and spontaneous, almost effortless | Broad range; does not have to restrict what they want to say | Clear, smoothly flowing, well-structured speech | Intelligible throughout; accent may be noticeable but does not affect it |
Source: CEFR Companion Volume (2020), Appendix 3, Qualitative features of spoken language (expanded with phonology). We left out accuracy and interaction to keep the table readable. Add accuracy back if grammar errors carry real risk in the job, such as written follow-ups to calls.
How to use the rubric without fooling yourself
Rate each quality separately, then take the level most qualities agree on. A candidate with C1 range and B1 fluency is not "C1 with a nervous moment"; they are probably B2 for practical purposes. Write down which recording drove each rating.
Use two listeners who rate independently before comparing notes. When they disagree by more than one level, the task was probably too short or the descriptor too vague for that person's answer. Listen again together, and fix the task for the next round.
What to leave out of the score. Leave out anything the CEFR does not treat as proficiency: accent itself, voice pitch, speaking speed beyond what fluency covers, confidence, and recording quality. A cheap phone microphone in a noisy room is not a language skill. If a recording is unclear for technical reasons, ask for another one rather than scoring it down.
How IELTS, TOEFL and Versant test speaking
The established tests take three quite different approaches, and each has a real case. The facts below come from each test owner's own pages, as of October 2026.
IELTS: a live examiner interview
IELTS Speaking is a face-to-face interview with a certificated examiner. It has 3 parts and lasts 11 to 14 minutes in total. Part 2 gives the candidate a task card, 1 minute to prepare, and about 2 minutes to speak. Examiners score fluency and coherence, lexical resource, grammatical range and accuracy, and pronunciation, each with equal weight, on the 0 to 9 band scale.
- Fluency and coherence25%
- Lexical resource25%
- Grammatical range and accuracy25%
- Pronunciation25%
TOEFL iBT: short recorded tasks
The TOEFL iBT Speaking section has 11 items in about 8 minutes. There are two task types: Listen and Repeat, and Take an Interview. ETS says test time can vary because the test adapts. Since 21 January 2026, TOEFL iBT reports each section on a 1 to 6 scale, which ETS describes as more intuitively aligned with the CEFR. Candidates take it at a test centre or at home with a live proctor.
Versant: automated scoring for employers
Pearson sells Versant tests to employers rather than to students. Its Speaking and Listening Test takes 17 minutes and, according to Pearson, assesses fluency, pronunciation and intelligibility alongside listening. Pearson states that all Versant scores are entirely AI-based, trained by expert graders, and that results arrive within minutes. Scores are reported on Pearson's Global Scale of English.
Side by side, and where a voice form fits
| IELTS Speaking | TOEFL iBT Speaking | Versant Speaking and Listening | Voice form screen | |
|---|---|---|---|---|
| Format | Live interview, 3 parts | Recorded tasks, adaptive | Recorded tasks | Recorded answers to your own prompts |
| Speaking time | 11 to 14 minutes | About 8 minutes | 17 minutes (with listening) | You choose, typically 6 to 8 |
| Scored by | Certificated examiners | ETS central scoring | AI, per Pearson | Your own trained listeners |
| Content | General and academic topics | General and academic tasks | General workplace English | Tasks from the actual job or course |
| Recognised outside your organisation | Yes | Yes | Recognised by employers who use it | No |
| Identity check and security | Test centre or proctored | Test centre or proctored at home | Proctoring options offered | None beyond your own process |
The last two rows are where the certified tests win outright, and no rubric closes that gap. A voice form's only advantage is the content row: you can ask people to do the exact speaking the job needs.
How to run a low-stakes speaking screen with a voice form
A low-stakes screen supports a decision a human still makes. Think of a shortlist for a support role, placing students into a conversation class, or checking a volunteer can brief a group. Nothing on a certificate depends on it.
- 1Decide the level the role or course needsPick one boundary to test, such as B1 or B2, from the real speaking tasks people will do.
- 2Write 3 or 4 tasks that climb in difficultyA familiar personal answer, a description, an explanation of a process, and an opinion with reasons.
- 3Set a time limit per answer and say it on screen1 to 2 minutes each, plus a short unscored warm-up question so nerves settle before rated tasks.
- 4Share the link, with no install or sign-inCandidates answer on their phone browser, one question at a time, and can re-record before moving on.
- 5Rate each recording against the rubricTwo listeners, independently, using fluency, range, coherence and intelligibility. Note the evidence for each rating.
- 6Treat the result as a band, not a scoreReport "likely B2" with notes, and send borderline cases to a live conversation or a certified test.
Tasks that produce ratable speech
Good tasks make candidates speak at length on something they know, then push slightly past comfort. Avoid trivia and anything that needs specialist knowledge, because then you rate knowledge instead of language. Four prompts that work for most work contexts:
- Tell us about your current or most recent job, and what a normal day looks like.
- Describe a place you know well to someone who has never been there.
- Explain how to do a task from your work, step by step, as if training a new colleague.
- Some companies let staff work from home full time. Do you think that is a good idea? Give your reasons.
Task 1 is the warm-up. Tasks 2 and 3 separate A2 from B1 well. Task 4 asks for a viewpoint with reasons, which is where B2 and above show. For more prompt ideas aimed at hiring, see our voice screening questions for hiring.
What Would You Rather does and does not do here
Would You Rather lets you record each prompt in your own voice. You can set minimum and maximum answer lengths per question. Candidates answer in a phone browser with no app. Results show each recording next to the candidate's other answers. A Google Sheets sync adds audio links you can rate in your own columns, as described in sending form responses to Google Sheets. It does not transcribe, score or level anyone. We think that is the right split for a low-stakes screen: software collects the samples, trained people judge them.
There is also a read-aloud mode built for speech-data projects, with transliteration suggestions for Indian languages. Reading a set script tests intelligibility on known words. It does not test fluency, range or coherence, so we would not use it as a speaking test on its own.
When you need a certified test instead
Some decisions need a result that people outside your organisation trust. In those cases a voice form, however well designed, is the wrong tool.
The result has legal or admission weight
Visa applications, university admissions and professional registration usually name specific accepted tests. Your rubric is not on those lists, and no amount of rigour puts it there. Check the receiving body's list first, and send candidates straight to a test on it.
You need to know who is speaking
A voice form cannot confirm the person recording is the applicant, or that nobody is prompting them. Certified tests handle this with test centres or live proctoring. If impersonation would be costly, use a proctored test or move the speaking check into a live video call.
You need comparable scores across thousands of people
At high volume, consistency across many raters gets hard, and an automated test like Versant becomes attractive. We would still ask any vendor for its validation evidence on speakers like yours. A 2020 PNAS study of five commercial speech recognisers found error rates of 0.35 for Black speakers against 0.19 for white speakers in US English. That paper tested transcription, not Pearson's scoring, but it shows why validation on your population matters.
What this approach does not tell you
The CEFR descriptors were written to describe ability across many contexts, not to rate a 6-minute recording. Placing someone at a level from 3 or 4 short answers is a judgement, and it is less reliable than a calibrated test. We have no reliability data for screens run this way, ours or anyone else's.
Untrained raters drift. Without a short calibration session, where both listeners rate the same sample recordings and discuss differences, two people can read "fairly even tempo" very differently. Certified tests invest heavily in examiner training for exactly this reason.
Our comparison table also simplifies. IELTS, TOEFL and Versant each publish technical and validation material we did not review here. Their alignment to CEFR levels is the test owners' own claim. The format and timing facts come from their public pages as of October 2026, and test formats change; TOEFL's changed in January 2026.
Finally, a recorded answer is not a conversation. Interaction, one of the six CEFR qualities, barely shows up when nobody talks back. If the job is mostly live conversation, finish with a short live call for anyone who passes the screen.
Start with one boundary and four prompts
Pick the single level boundary your decision really turns on. Use the four prompts above, record them in your own voice, and set 1 to 2 minutes per answer. Before you rate real candidates, have two colleagues rate 5 sample recordings and agree on what B1 and B2 sound like for your context.
If your candidates speak several languages, read collecting answers in Hindi, Tamil or any language and how code-switching shows up in voice answers before you launch. If you are assessing students rather than candidates, oral assignments online covers the classroom version of the same idea.
References
- Council of Europe, Common European Framework of Reference for Languages: Learning, Teaching, Assessment. Companion Volume, 2020. Used for the rubric descriptors, the phonological control rationale and the native-speaker statement. Read from an archived copy because the official host blocked automated access.
- IELTS, IELTS Academic Speaking format and Scoring in detail, opened 8 October 2026. Used for Speaking format, timing, criteria and weighting.
- ETS, TOEFL iBT test content and Understanding your TOEFL iBT scores, opened 8 October 2026. Used for Speaking tasks, timing and the 1 to 6 scale.
- Pearson, Versant by Pearson English tests for HR professionals, opened 8 October 2026. Used for Versant test length, AI scoring and reporting scale.
- Koenecke, A. and colleagues, Racial Disparities in Automated Speech Recognition, PNAS 117(14), 2020. Used for speech recognition error rates by speaker group.
The weakest point in this source base: all test facts come from the test owners' own pages. We also found no independent study of short, form-based speaking screens like the one we recommend.
Questions people ask
How can I test my spoken English online?
For a recognised result, book a certified test such as IELTS, TOEFL iBT or a Pearson Versant test; all report scores that institutions and employers recognise. For practice or an informal check, record yourself answering a few open questions for 1 to 2 minutes each, then compare your recordings with the CEFR descriptors for fluency, range, coherence and intelligibility. A teacher or colleague listening against the same descriptors gives a more reliable read than self-rating.
What is a language proficiency assessment?
A language proficiency assessment measures what someone can do in a language now, regardless of how they learned it. It differs from an achievement test, which checks what was taught in a specific course. Most proficiency assessments report against a shared scale, commonly the CEFR levels A1 to C2, so results can be compared across tests, employers and countries. Speaking is usually assessed separately from reading, writing and listening.
Which spoken English test is accepted for jobs?
Employers set their own rules. Many accept IELTS or TOEFL iBT results, and Pearson markets its Versant tests directly to employers for hiring, with automated scoring and results within minutes. For regulated professions or visa sponsorship, check the regulator's or immigration authority's list of approved tests, because only named tests count there. For internal screening, an employer can use any method that is consistent, job-related and applied equally.
What CEFR level is good spoken English for work?
It depends on the job. B1 users can handle familiar topics and simple conversations, with evident pauses for planning. B2 users can give clear descriptions and express viewpoints on most general topics with a fairly even tempo, which many customer-facing roles need. C1 users speak fluently and spontaneously on complex subjects. Define the level from the tasks the person will actually do, not from a general preference for polish.
Can AI score a speaking test accurately?
Automated scoring is now standard in some certified tests; Pearson states its Versant scores are entirely AI-based, trained by expert graders. Accuracy for your candidates depends on how well the system was validated on similar speakers. Independent research has found commercial speech recognisers make more errors for some groups of speakers. For a do-it-yourself screen, have trained humans listen and score against the rubric.
How long should an online speaking test be?
Long enough to hear several types of speech, short enough that people finish it. Certified tests use about 8 to 17 minutes of test time for speaking. For a low-stakes screen, 3 or 4 tasks of 1 to 2 minutes each, with a short warm-up first, gives a listener enough to place someone within a CEFR band. Longer answers rarely change the rating, and they add listening time for every reviewer.
Drafted with AI from published research, then edited and approved by Shivam Singh.
Would you rather hear the answer?
Make a voice form in a minute. It’s free.
Create a form