Research technology · working prototype
Predict the next listening miss.
EarShift is developing a learner model that uses a short test to predict which known English a person will miss in new natural speech—and tests whether targeted practice transfers beyond the trained recording.
Try the public demo75-second product demo
Measure. Explain. Train. Test transfer.
The video shows the working prototype and labels planned research separately. Independent vocabulary checks and systematic new-speaker validation are next steps; improved film comprehension is not yet established.
Eight-slide common deck
The research thesis, early evidence and Phase I plan.
The deck frames EarShift as a high-risk measurement and prediction technology—not a collection of language-learning features.
Founder
Yuliia Sereda
Founder · Senior Software Engineer
After recently moving to the United States, Yuliia discovered that knowing English vocabulary and grammar did not guarantee understanding fast, connected everyday speech. English had become the language of daily life: she needed to understand people around her and speak clearly enough to be easily understood. The same gap followed her into films and online video, where she often missed important details without subtitles.
While looking for ways to improve both listening and pronunciation, she reached a simple insight: it is harder to reproduce sound distinctions and rhythm that are not yet reliably heard. Research on targeted perceptual training suggests that specific L2 listening distinctions can improve, although transfer to unfamiliar natural speech and production is neither complete nor automatic. EarShift turns that boundary into a testable question.
Yuliia has nine years in software engineering, B.S. and M.S. degrees in Computer Science, and experience building backend services for 50,000+ concurrent users and 300,000+ daily requests. She leads EarShift’s research prototype, data architecture and formative studies.
Evidence boundary
Feasibility demonstrated. Efficacy not yet demonstrated.
Eight compensated participants completed two formative rounds. The latest three-person pilot had complete core event capture and 4.7/5 average engagement, while also exposing ceiling items and parallel-form imbalance. These results justify the next controlled study; they do not establish learning impact.
- 01Prediction
Test whether sparse diagnostic observations predict errors on held-out recordings, lexical contexts and speakers.
- 02Causal source
Separate vocabulary unfamiliarity from known-but-missed speech using independent lexical probes.
- 03Transfer
Compare adaptive selection with an equal-time active control on immediate and 24–72-hour delayed unseen speech.
Experience the instrument as a learner.
The public demo takes about 25–30 minutes and works best with headphones.
Open the listening lab