유대선
← 프로젝트로
·업데이트·2 분

Realtime voice sparring v1 — your due SRS cards ride into a live conversation

First in-app realtime voice feature: OpenAI Realtime sparring where today's due SRS cards are secretly planted in the conversation, detected when you actually say them, and graded on the spot. Shipped with zero new native dependencies via a hidden WebView WebRTC bridge.

MiMi has drills, shadowing, and a Leitner SRS — but everything ended at "외웠다". Today the loop closes: a realtime voice sparring partner (OpenAI gpt-realtime, native speech-to-speech) where the conversation itself is the review.

How it works

  1. The app picks up to 6 of today's due SRS cards (verb pack, phrasal-500, collocations — short, sayable chunks only).
  2. The backend mints an ephemeral OpenAI Realtime secret and builds the session instructions: a fixed persona prefix (prompt-cache friendly) plus a secret coaching goal — steer the conversation so the learner gets natural openings to use those exact expressions, without ever quizzing.
  3. Audio flows phone ↔ OpenAI directly over WebRTC. When the live transcript shows the learner actually said a target ("figured it out" matches "figure out" — conjugation-tolerant, separable-phrasal-verb-tolerant), the chip turns green and the card is graded correct through the existing SRS API.
  4. Two modes, same engine, different 사용감: Free talk (250 ms VAD — a quick-witted friend who pounces the moment you finish) and Mock interview (800 ms VAD — a patient Canadian tech interviewer who lets an ESL candidate organize long answers). The interview mode is the founder's own Canada job prep, dogfooded.

The interesting engineering choice

No react-native-webrtc. The app already ships react-native-webview (YouTube transcripts), so v1 borrows the browser's WebRTC through a hidden, headless bridge page served from the backend over HTTPS: React Native injects the ephemeral secret, the page handles mic ↔ OpenAI audio, and transcripts come back over postMessage while all visible UI stays native. Zero new native dependencies, zero new build risk, and the transport can be swapped for a native module later without touching the product logic.

Key custody stays clean: the raw OPENAI_API_KEY never leaves the server (it was already there for transcription); the client only ever holds a single-session, short-TTL secret.

Verified

Commit: 048c02e.