AI IELTS Prep
Speak into it and get a band score back with reasons. A speaking examiner built in 48 hours at a hackathon.
The situation
Speaking practice is the one part of IELTS you cannot do alone. You need someone to talk to who also knows the band descriptors, and that person is expensive, scarce, and not available the night before your test.
The hackathon brief was open, so I scoped a preparation platform across all four sections, listening, reading, writing and speaking, matched to the real exam. Forty questions in thirty minutes for listening, sixty minutes for reading, two writing tasks, and a three part speaking interview of eleven to fourteen minutes. The speaking part was the interesting engineering.
How it works
The diagram is the pipeline. No recording survives, so the steps are described and the rubric is shown.
- Voice activity detectionDetects when the candidate starts and stops talking so the recording does not need a button.
- TranscribeWhisper turns the answer into text.
- StoreSave the turnThe conversation is stored so follow up questions have context.
- GateScoreMistral grades the answer against fluency, vocabulary, grammar and pronunciation and drafts the examiner’s next question.
- RespondText to speech reads the feedback and the next question back to the candidate.
- Other sectionsListening, reading and writing use simpler endpoints that take the context, the question and the answer and return feedback.
Answer the examiner
Voice activity detection starts the recording when you speak and stops it when you pause, the way a real interview does.
Get transcribed and scored
Whisper transcribes the turn and Mistral scores it against the four speaking criteria with a band from one to nine and a reason for each.
Hear the next question
The examiner’s follow up comes back as speech, so the practice keeps the rhythm of the real test.
Practise the rest
Listening, reading and writing each have an evaluation endpoint that grades an answer against its question, with question types matched to the exam.
Fluency
- Effort and ease of speech
- Hesitations and self correction
- Connectives and discourse markers
- Ability to develop a topic
Vocabulary
- Range of words used
- Accuracy of word choice
- Idiomatic use and collocation
- Paraphrase when a word is missing
Grammar
- Simple and complex structures
- Flexible use of tenses
- Mistakes and their effect on understanding
Pronunciation
- Range of pronunciation features
- Control and consistency
- Ease of understanding
Decisions that mattered
Scope all four sections, build the hard one properly
The platform matched the real exam structure so the demo felt complete, but the engineering effort went into the speaking loop where the interesting problems were.
Detect speech instead of using a button
Voice activity detection makes the interview feel like a conversation and removes the awkward start and stop.
Score against the real descriptors
The prompt uses the published IELTS speaking criteria so the feedback maps to what an examiner would say, not a generic rating.
Where it landed
A browser based speaking examiner that listens, transcribes, scores and answers back, with the other three sections scaffolded to the real exam. It won AI Crusade 2024.
- Speak, transcribe, score, respond loop working end to end
- All four sections matched to real exam timings
- Won AI Crusade 2024
- Only the speaking rubric was fully written. The other three sections had endpoints but thin scoring guidance.
- It was never production grade after 48 hours and does not claim to be.
- No recording of the build survives, so this page uses a diagram and the rubric instead of stills.
Notes from the build
Written in my own words, after the fact. Opinions, not claims.
Hackathons taught me to find the one loop that has to work and build everything else as scaffolding around it. Here the loop was speak, transcribe, score, respond. Once it ran end to end the rest was forms.
The scoring guidelines file from that weekend is still the clearest thing I have written about what a rubric needs to say to a model. Range, accuracy, and the effect of mistakes on understanding, in that order.
Nawap Bastola · 48 hours in 2024