SpeakMetry

Most IELTS Speaking prep is either a tutor you have to book, or a stack of practice books you read silently. Neither one puts you in the actual conditions of the test: speaking out loud, under time pressure, to something that responds back. That gap is why I built SpeakMetry.

The core of it is a real-time voice AI examiner running on OpenAI's Realtime API, driven by a state machine that tracks exam phase, Part 1 warm-up questions, Part 2's one-minute prep plus timed long turn, Part 3 follow-up discussion, the same structure and timing as the real exam. After each session, GPT-4o scores the response against the actual IELTS band descriptors (fluency and coherence, lexical resource, grammar, pronunciation) instead of spitting out one vague number.

The hardest part wasn't the AI, it was the state machine. Getting the exam to feel authentic meant handling timing precisely, the one-minute Part 2 prep has to actually feel like one minute, transitions between parts can't have dead air, and the examiner has to interrupt naturally at time limits the way a real examiner does, not abruptly cut off mid-sentence. That's mostly been a UX/engineering problem, not a model problem.

On pricing: I deliberately didn't make this a subscription. Everything is either free with daily limits, or a one-time Premium Pass for a fixed prep window (1/3/6/12 months, never auto-renews), plus pay-as-you-go credits for the realistic AI exam sessions specifically, since those cost more to run than the free daily practice tools. Most people prep for IELTS for a few weeks to a couple months, not indefinitely, so a recurring subscription felt like the wrong model for how people actually use it.

Built solo, still actively shipping. If you've prepped for IELTS or a similar speaking exam, I'd genuinely like to know: what part of the format is hardest to practice alone? That's mostly what's shaping what I build next.

speakmetry.com if you want to try it.