How CELPass Scores Your CELPIP Speaking: Transcript, Measured Delivery and Five Criteria
By David · Speaking · Published · 12 min read
Speaking is the hardest CELPIP skill to practise alone: you cannot hear your own pauses, and a friend cannot tell you whether an answer is CLB 7 or CLB 9. CELPass scores every Speaking answer with two engines working together: audio analysis, which listens to your recording and measures how you speak, and a large language model, which judges what you said. SpeechTrace™ turns the audio into measurements of pace, pauses, fillers and clarity, and BandBridge™, our scoring engine, rates five criteria with quoted evidence for each. This post goes through every step, the rules that keep the score honest, and what you get back besides a number.
How does CELPass score a CELPIP Speaking answer?
It analyses the audio of your recording to get a timed transcript and measurements of your delivery, then has a large language model score five criteria against CELPIP-style level descriptions, quoting your own words as evidence. The overall level is the rounded average of those criteria, lowered only when the engine can point to a weakness that clearly gets in the way of communication.
- Audio analysis: a deep-learning speech recognition model listens to your recording and returns every word with its start time, end time and a clarity score.
- Delivery measurement: SpeechTrace™ computes your speaking rate, long pauses, hesitation fillers and word clarity from that audio analysis. This step is plain arithmetic, so the same recording always gets the same numbers.
- Language rating: BandBridge™ gives a large language model the task prompt, your transcript and the delivery measurements, and it scores each criterion with 2 to 4 observations from your answer.
- Final level: the criterion scores are averaged and rounded to a whole level; a cap is applied only with cited evidence, and the reason is shown to you.
- Feedback: your transcript comes back with mistakes marked, the corrections and rules, a higher-band version of your answer, and drills built from your own sentences.
Why does CELPass use both audio analysis and a language model?
Because a Speaking answer has two sides, and no single model judges both well. A language model is good at judging ideas, organisation, vocabulary and grammar, but it only reads text, so it cannot hear a two-second silence or a string of fillers. Audio analysis hears exactly that, but cannot tell whether your advice made sense. CELPass gives each job to the engine built for it, then combines the results.
| Engine | What it works on | What it judges |
|---|---|---|
| Audio analysis (speech recognition + SpeechTrace™) | Your recording | Speaking rate, long pauses and when they happen, hesitation fillers, how clearly each word came through |
| Large language model (BandBridge™) | The transcript, the task prompt and the audio measurements | Content/Coherence, Vocabulary, Grammar, Task Fulfillment, and Fluency & Pronunciation from the measured delivery |
The two meet in Fluency & Pronunciation, which is scored only from the audio measurements and the hesitations in your transcript. A text-only scorer has to guess this criterion; CELPass measures it, and when a recording has no usable audio data it leaves the criterion blank instead of guessing.
The size of the model matters less than how it is used. A big model given a vague instruction like "rate this answer" tends to be generous. BandBridge™ gives the model written level descriptions that work as ceilings, anti-bias rules, and a required format: a score, quoted evidence and what is missing for the next level, for every criterion. That structure is what keeps scores from drifting upwards.
Why does CELPass measure delivery instead of judging it from the transcript?
Because a transcript cannot show how you sounded. Two answers with the same words can be a smooth 60 seconds or a halting one full of long silences, and a text-only scorer would give them the same rating. CELPass keeps the timing of every word, so pace and pauses are counted rather than guessed.
| SpeechTrace™ measurement | How it is calculated |
|---|---|
| Speaking rate | Words per minute over your actual speaking time, from your first word to your last, so a quiet start does not count against you |
| Long pauses | Every silent gap of 2 seconds or more between two words, with the moment it happened |
| Hesitation fillers | Each um, uh, er, ah, hmm or mm the recogniser picked up |
| Word clarity | How confidently the recogniser understood each word, the share of words it was unsure about, and which ones |
As a rough reference, a CLB 9 answer in our scoring runs at about 120 to 150 words per minute, with fewer than 3 fillers and at most 1 long pause. That is a guide for the rater, not a hard rule: a slightly slower answer with well-built ideas still scores well, and speed alone is never rewarded.
Word clarity is handled with care. It drops when the recogniser is less used to an accent, but also with background noise, a weak microphone or an unusual name. So it counts only as an indirect hint about pronunciation, never as the only reason for a lower score, and the engine is told not to penalise accent.
Which criteria does CELPass score in Speaking?
Five, each on a level from 2 to 12: Content/Coherence, Vocabulary, Grammar, Task Fulfillment, and Fluency & Pronunciation. They follow the four dimensions CELPIP raters use, with Listenability split into its grammar part and its delivery part so you can see which one is holding you back.
| CELPass criterion | Official CELPIP dimension | What it looks at |
|---|---|---|
| Content/Coherence | Content/Coherence | Whether your ideas are relevant, developed with reasons and examples, and in a logical order |
| Vocabulary | Vocabulary | Range, precision and natural word choice for the topic |
| Grammar | Listenability | Sentence range and control, and whether errors get in the way of meaning |
| Fluency & Pronunciation | Listenability | Pace, pauses, fillers and clarity, judged from the SpeechTrace™ measurements |
| Task Fulfillment | Task Fulfillment | Whether you did what the task asked, for the right listener, with a suitable tone, in the time |
If a recording has no usable word timings, Fluency & Pronunciation is left blank rather than guessed, and the overall level is the average of the other four. You see a missing score instead of an invented one.
What does each Speaking level look like to the scoring engine?
Each level has a written description, and the descriptions work as ceilings: an answer has to show the features of a level to receive it. The most important line sits between CLB 9 and CLB 10, because many answers that are clear and accurate still stop at 9.
| Level | What the engine expects |
|---|---|
| 11-12 | Confident, precise spoken English; well-developed ideas with depth; flexible vocabulary; errors rare and insignificant |
| 10 | Consistently beyond "clear and coherent": effective elaboration, good range of words and structures, errors minor and infrequent |
| 9 | Clear and coherent throughout; minor errors and occasional awkward phrasing; ideas relevant but may lack depth |
| 8 | Clear overall message with noticeable errors or simplifications; logical but basic organisation; limited vocabulary range |
| 7 | Meaning clear despite frequent errors; basic organisation; simple but adequate vocabulary; task completed |
| Below 7 | Meaning often unclear, organisation breaks down, the task is only partly done, or errors cause confusion |
The ceiling rule: an answer that is clear, coherent and accurate but shows limited development, vocabulary variety or grammatical range does not go above 9. To reach 10, the answer has to show more, such as a reason followed by a specific example, precise words instead of general ones, and a mix of sentence structures used correctly.
When is a CELPass Speaking score capped?
Only when the engine can quote part of your transcript showing that one weakness clearly interferes with clarity, coherence or task completion. When that happens, the results page says the score was capped and explains why. There are no automatic caps that fire on a single feature.
Some scorers use fixed caps, such as a set maximum for any answer under a time limit. We chose not to, because the same feature matters differently from task to task: a 50-second answer to a 60-second task can be complete, while a 70-second answer to a 90-second task that skips half of the prompt is not. Instead, short and off-topic answers lose points where the problem actually is, in Content/Coherence and Task Fulfillment, and a cap is added only when the evidence shows the whole answer is affected.
- An answer that drifts off the topic: Task Fulfillment and Content/Coherence drop, and a cap is likely because the listener does not get what the task asked for.
- An answer that stops early: SpeechTrace™ reports your speaking time and word count, and the rater judges whether the task was still completed.
- One weak criterion on an otherwise strong answer: it lowers the average, but does not cap the level unless it gets in the way of the message.
| Task | Preparation | Speaking time |
|---|---|---|
| 1. Giving Advice | 30 seconds | 90 seconds |
| 2. Talking about a Personal Experience | 30 seconds | 60 seconds |
| 3. Describing a Scene | 30 seconds | 60 seconds |
| 4. Making Predictions | 30 seconds | 60 seconds |
| 5. Comparing and Persuading | 60 seconds | 60 seconds |
| 6. Dealing with a Difficult Situation | 60 seconds | 60 seconds |
| 7. Expressing Opinions | 30 seconds | 90 seconds |
| 8. Describing an Unusual Situation | 30 seconds | 60 seconds |
How does CELPass keep Speaking scores fair?
The engine follows written anti-bias rules for spoken English. It judges only what is in your answer, tolerates the normal features of speech, and lets an error lower the score only when it gets in the way of meaning, coherence or the task.
- Evidence only: the engine does not assume ability you did not show, good or bad.
- No default low scores: an answer that is understandable and completes the task starts from CLB 7 and is not pushed below it unless clarity, coherence or task completion clearly breaks down.
- Speech is not writing: a self-correction, a repeated word, an occasional filler or informal phrasing is normal when speaking and is not punished on its own.
- No accent penalty: accent is not a criterion. Delivery is judged from measured pace, pauses and fillers.
- No reward for speed alone: fast, fluent speech with thin ideas does not score high.
The rules work in both directions. A flattering score is as unhelpful as a harsh one: it can make you book the test too early. That is why every criterion comes with the reason it is not yet at the next level, so a 9 tells you what a 10 would need.
What feedback do you get after a Speaking answer?
Besides the level, you get your recording and transcript side by side, every flagged mistake with its correction and rule, a short examiner-style summary, and a higher-band version of your own answer to compare with.
| Part of the results | What it shows |
|---|---|
| Scores | Your overall level and each of the five criteria, with 2 to 4 observations each and what is missing for the next level |
| Transcript | What the recogniser heard, with fillers, grammar, vocabulary and coherence mistakes marked in the text |
| Fix these next time | Up to 6 of the most important mistakes: your exact words, the correction and a one-line rule |
| Delivery | Your speaking rate, long pauses with their timing, filler count and the words the recogniser found unclear |
| A higher-band version | Your answer rewritten with the same ideas, but clearer, better developed and better organised |
| Try it again tomorrow | A prompt to record the same task again, so you can compare the two attempts |
The higher-band version keeps your ideas on purpose. Comparing it with your own answer shows exactly which phrases and links changed, which is easier to learn from than a model answer on a different idea.
What happens to your Speaking mistakes after scoring?
Each flagged mistake is filed under a fixed pattern, such as articles, verb tense, missing linking words, a weak conclusion or long pauses, and becomes a drill card built from your own sentence. Cards come back on a spaced-repetition schedule, and better words from your answers are saved to your Words list.
Because the patterns are fixed, CELPass can count them across all your answers. After a few attempts you see which two or three patterns cost you the most, which is usually a faster way up than practising every task equally.
Why does CELPass show a CLB range for Speaking?
Because one answer is not enough evidence for an exact level. Each answer gets a whole-level score, but your predicted Speaking CLB comes from your recent scored attempts and is shown as a range that narrows as you practise more tasks.
Speaking scores vary more than most learners expect: a familiar topic in Task 1 can score a level higher than a picture you have never seen in Task 3. A range built from several tasks is a much better guide to test day than your best single answer.
Is a CELPass Speaking score the same as an official CELPIP score?
No. It is a practice estimate on the same dimensions CELPIP raters use, built to avoid over-scoring. Only the official CELPIP test, rated by Paragon's trained raters, gives a score you can use for immigration or citizenship. CELPass is independent and not affiliated with Paragon Testing Enterprises or IRCC.
Use it to find the criterion holding you back and to check that you are moving in the right direction over several attempts. If your range sits at or above your target across different tasks and a full mock test, that is a good sign you are ready to book. If a score looks wrong to you, tap 👎 under it and tell us why: we read every report.
Frequently asked questions
Does CELPass penalise my accent?
No. Accent is not a criterion, and the engine is told not to penalise it. Delivery is judged from measured pace, pauses and fillers; word clarity is only an indirect hint and never the only reason for a lower score.
Do fillers like "um" lower my Speaking score?
An occasional filler is normal and is not punished on its own. Many fillers, or long silences, lower Fluency & Pronunciation because they make you harder to follow.
Why is my Speaking score capped?
A cap is applied only when the engine can quote part of your answer showing one weakness clearly interferes with clarity, coherence or the task, such as answering a different question. The results page shows the reason.
Can background noise affect my CELPass Speaking score?
It can lower the word-clarity measurement and make the transcript less accurate, so record in a quiet room with a decent microphone. Clarity alone never decides your score.
How many Speaking answers do I need for a reliable estimate?
One answer gives a score for that answer. For a reliable Speaking CLB, answer several different tasks or take a full mock; the predicted range narrows as you add attempts.
Keep reading
CELPass is an independent practice platform and is not affiliated with Paragon Testing Enterprises. CELPIP is a trademark of Paragon Testing Enterprises. Check the official CELPIP website and IRCC for current test rules and requirements.