How does Britwise score IELTS Speaking?

Britwise scores based on 4 criteria from Cambridge Public Band Descriptors: Fluency & Coherence, Lexical Resource, Grammatical Range & Accuracy, Pronunciation. AI uses OpenAI Whisper for voice conversion and GPT-4o for scoring, along with detailed Examiner Notes.

How many mock tests are available on Britwise?

Currently, there are 117+ Cambridge-style mock tests, plus new tests generated weekly according to Cambridge style and rubric — you'll never run out of practice materials.

How does Britwise's money-back guarantee work?

100% refund within the first 7 days from the date of payment. Additionally, a +0.5 band guarantee within 30 days if the learning path is completed — if not, a full refund of the first month's package.

How is Britwise different from ELSA or other apps?

Britwise focuses 100% on IELTS Speaking with the Cambridge-Sync™ method, authentic British English voice, 14-minute mock tests closely aligned with the exam, and detailed Examiner Notes. This is an in-depth IELTS coach, not a general pronunciation app.

Does Britwise support local payment methods?

Yes. Britwise supports local payment methods and credit cards. VAT invoices can be requested via support@britwise.school.

Efficacy & Benchmark — Britwise School

EFFICACY · METHODOLOGY · PROOF

How accurate is the Britwise examiner?

Most AI English-grading vendors quote a single accuracy number and refuse to publish their methodology. We publish the rubric, the models, the dataset, the gaps, and the human-examiner panel that grades against us. If it disagrees with us, we publish that too.

Cambridge-aligned rubric

Human examiner reference panel

External audit pending

OUR STACK

The models that grade your test

No black boxes. Every component, every model, every latency budget.

Task

Model

Latency

Notes

Speech-to-text

Deepgram Nova-2

~600 ms

EU Frankfurt endpoint · zero-retention · filler-word capture

Pronunciation prosody

Hume AI

~400 ms

Confidence + emotion + cadence — feeds into Pronunciation band

Speaking grading

OpenAI GPT-5.2

~2.2 s

Cambridge 4-criteria rubric · temp 0.2 · structured JSON

Full mock exam grading

OpenAI GPT-5.5

~3.0 s

Reasoning-heavy · used for end-of-week mock tests

Writing Task 1 / 2

OpenAI GPT-5.2

~2.5 s

Cambridge 4-criteria · 1500-token feedback + B-level rewrite

Reading / Listening MCQ

Deterministic key

<50 ms

Cambridge mock-test answer keys · band-scaled per official conversion

Coach voice (Angie)

ElevenLabs Flash v2.5

~75 ms first byte

British RP · stream-first

All LLM calls route through the Britwise abstraction layer — we can swap providers without breaking the rubric. Audio is processed in the EEA where possible; no audio is retained by Deepgram or OpenAI (zero-retention contracts on file).

ACCURACY

How close are we to a human Cambridge examiner?

The honest table. Where we’re using extrapolated numbers (because the 100-sample study is still in flight) we say so.

Metric

Human examiner

Britwise (GPT-5.2)

Typical AI competitor

Status

Within ±0.5 of human (target)

70%

≈ 70% (extrapolated)

≈ 60%

study in progress

Within ±1.0 of human

95%

≈ 92% (extrapolated)

≈ 88%

study in progress

Off by ≥ 2.0 bands (catastrophic)

<1%

≈ 1.6% (extrapolated)

≈ 5%

study in progress

Inter-attempt consistency (same sample, 5 runs)

n/a

0.32 σ

0.61 σ

internal test, June 2026

Reference numbers for “human examiner” come from Cambridge ESOL inter-rater reliability studies (Taylor & Galaczi, 2011; Cambridge Research Notes vol. 65). Competitor estimates are extrapolated from published GPT-4 IELTS papers (2024) since most vendors do not publish their own numbers.

METHODOLOGY

Five steps. No theatre.

The exact procedure we follow to produce the numbers above.

01

Sample

100 anonymised candidate audio submissions from the past 60 days — stratified across Bands 4 → 8 (20 per band). Personal identifiers stripped; consent recorded under the Britwise Privacy Notice §4.

02

Reference panel

Each sample is independently graded by THREE Cambridge-certified examiners (recruited via the British Council network) using the public IELTS Speaking Band Descriptors. The reference band is the median of the three.

03

System grade

Britwise grades each sample 5 times using GPT-5.2 with temperature 0.2 and our production rubric prompt. The Britwise band is the mean of the 5 runs (rounded to the nearest 0.5).

04

Metrics

We report: % within ±0.5 of reference · % within ±1.0 · catastrophic disagreement rate · run-to-run standard deviation. All raw data published on this page in the JSON download.

05

Audit

An external party (target: a UK NCFE-registered awarding body) audits the dataset and methodology. The audit report is published in full once received.

KNOWN GAPS

Where we’re honest about the limits

If a vendor claims 99% accuracy without publishing their dataset, run.

Empirical benchmark in progress — numbers above are extrapolated from published GPT-5 family papers, not from our own dataset yet.

Pronunciation grading currently leans on Hume AI prosody + LLM judgement, not yet phoneme-level scoring (planned: integrate Microsoft Azure Speech Pronunciation Assessment).

Inter-rater reliability between human Cambridge examiners is itself only ~70% within ±0.5 — no AI system can exceed that ceiling.

Speaking-Part-2 long-turn (the 2-minute monologue) is the hardest task; our error is concentrated here. We expect to publish Part-1/2/3 split numbers in the next study.

FOR PROCUREMENT TEAMS

Need the dataset, the rubric, or to speak with an examiner on the panel?

Email efficacy@britwise.school for the full methodology PDF, signed examiner CVs, and access to the redacted candidate audio used in the study. Available under NDA for active enterprise procurement.

Talk to procurement

Trust centre →

Privacy · Terms · Cookies · DPA · SCC · DPIA · Sub-processors · Responsible AI · Accessibility · Security · Modern slavery · Complaints

Britwise School LTD

Company No. 17253094 · ICO ZC174279 · Registered in England & Wales 71–75 Shelton Street, Covent Garden, London, WC2H 9JQ, United Kingdom

🇬🇧 British English · Cambridge-grade scoring · GDPR compliant