How scoring works
What NOVA measures, what it estimates, and what it refuses to claim.
A practice score is only useful if you know how it was made. This page is the whole method, written from the code that runs it. Where something is an estimate, it says estimate. Where nothing is measured, it says so.
The pipeline
Six steps between your voice and a number.
- 1
You speak
Under the timing of the part you chose. Everything is transcribed so you can read what was heard.
- 2
The evidence is checked
Is there enough speech? Is the audio usable? The answer sets the score basis on the report.
- 3
Measured in code
Pace, silence, repeats, connectors and lexical diversity are counted, with no model involved.
- 4
Judged by the model
A bounded rubric, one band per criterion in half steps, every finding tied to a quote.
- 5
Checked before you see it
Quotes verified, claims filtered, missing evidence marked not assessed.
- 6
Labelled and stored
Score basis, model, prompt, rubric and calibration versions are saved with the report.
Criterion by criterion
Measured, or judged. Never blurred.
Some signals are counted by code and are the same every time. The bands themselves are judged by a language model against a bounded rubric. Here is which is which.
Fluency and coherence
MEASURED IN CODE
- Words per minute, when the recording has usable timing
- Share of silence, when the recording has usable timing
- Fillers and immediate word repeats per 100 words
- Connectors per 100 words, and how many different ones
JUDGED BY THE MODEL
- A band from 4 to 9 in half steps
- One strength and up to two priorities, each tied to a quote from your transcript
Lexical resource
MEASURED IN CODE
- Lexical diversity as a moving-average type–token ratio. Teachers see the number; learners see a level once it is calibrated
JUDGED BY THE MODEL
- A band from 4 to 9 in half steps
- Findings on range, precision, collocation and paraphrase, with quotes
Grammatical range and accuracy
MEASURED IN CODE
- Nothing is measured mechanically for this criterion today. We say so rather than invent a metric
JUDGED BY THE MODEL
- A band from 4 to 9 in half steps
- Findings on range, accuracy and error patterns, with quotes. Recurring patterns become review cards
Pronunciation
MEASURED IN CODE
- Only from audio, never from a transcript. A recording qualifies if it is long enough, loud enough, mostly speech and not clipped
- An open-source acoustic model scores accuracy, fluency and prosody on its own 0–10 scale. That is not a band and is never shown as one
JUDGED BY THE MODEL
- An AI-estimated IELTS-style band from the recording itself, needing at least 15 seconds of active speech. The model listens three times and the middle result is kept, so a retry cannot move it
The pronunciation band is labelled an AI estimate, shown separately, and has not been calibrated against examiner ratings.
Speech measures need at least 20 words. The rubric is YNova's own short paraphrase of the public band descriptors, not the official text, and it is never presented as an official score.
Before you see a number
Six checks every report has to pass.
The model's output is not trusted. It is verified against your transcript and a set of rules, and if it fails, you get less, not a guess.
- Every quoted example must appear word for word in your transcript. A quote the model made up fails the report.
- No pronunciation band without qualified audio. A transcript alone cannot score pronunciation.
- No band numbers hidden inside the written feedback, and no wording that claims to be official or from an examiner.
- No explanation may be based on your first language, nationality or background.
- Recommended next steps come only from an approved list of things the app can actually do.
- If a check fails, the report is repaired once. If it still fails, you get a partial report with criteria marked not assessed, never a guessed number.
The overall band
Computed by code, not by the model.
The model never produces an overall band. It cannot; there is no field for one. The overall is computed afterwards by fixed rules, in half steps.
- Practice reports show an overall band only when all four criteria could be assessed. Otherwise there is no overall band.
- Mock test headlines average fluency, lexical resource and grammar. When an AI pronunciation estimate exists it is included, and the report says so.
- The averaging scheme used is saved with each attempt, so an old report is never silently rescored by a new rule.
- A personal best is only claimed against reports scored with the same rubric, calibration and evidence basis. The app never claims you got worse.
Score basis
Every report says what it was scored from.
- Transcript only
- Text criteria judged from the transcript. Pronunciation not assessed.
- Transcript and timing
- As above, plus pace and silence measured from the recording's timing.
- Audio and transcript
- Qualified audio exists. Pronunciation can be estimated and is labelled as an AI estimate.
- Not scored
- Too little evidence. You get feedback, not a number.
Practice estimate. Not an official IELTS result.
What we do not claim
Four sentences we will not write until we can.
“NOVA's bands match an examiner's.”
The scoring is a prompted baseline. The calibration study against teacher and examiner marks has not run. Until it has, the bands are labelled estimates and nothing more.
“Your score is exact.”
The judging model does not run at a fixed temperature, so a re-run can differ slightly. Every report keeps its input hash, model and prompt version so any score can be replayed and investigated.
“This is your IELTS result.”
Nothing in the app is an official IELTS score. The truth label sits next to every band, in every mode, and the report engine is forbidden from claiming otherwise.
“Pronunciation is scored like the rest.”
It is estimated from audio by a separate path, shown separately, and marked as an AI estimate. A transcript alone never produces a pronunciation band.
The studies that would earn these claims are listed on the research page. The models in use: Google Gemini for conversation and judging, ElevenLabs for speech to text and text to speech, and a self-hosted open-source acoustic model for pronunciation. Production reports are never generated by preview models.
Questions this page did not answer?
The FAQ covers accuracy, hallucination, privacy, what teachers see and why not just use a chatbot.