Problem
Shadowing apps reuse YouTube caption cues, which cut across sentence boundaries and break in awkward places.
Approach
We take word-level auto-captions and let an LLM split on grammatical boundaries. Measured across four videos, awkward breaks ran 38-44% for the benchmark app and 2-6% for ours.
What I built
- The full study loop: A-B repeat, merge and split, and a hide mode
- Azure pronunciation assessment gives per-sentence scores, word-level coloring, and Korean coaching
- Live on both the App Store and Play Store alongside the web app
- Full recurring-billing cycle verified on PortOne V2 with KakaoPay billing keys
- Only the caption extraction that serverless cannot do runs on an Oracle VPS worker
Stack
Next.js 16React NativeSupabaseVercel AI GatewayAzure SpeechPortOne V2yt-dlp
Next