← ClaudeAtlas

kokoro-g2-transcription-looplisted

Generate a deterministic Kokoro TTS phrase on a local computer, play it through the computer speakers for Even Realities G2 glasses to capture, collect Work Bench Android logs, score the resulting transcript, and diagnose transport, audio, VAD, transcription-worker, disconnect, and recovery failures. Use when validating or debugging the Work Bench G2 audio-to-transcript path on a physical Android device.
aaronrau/workbench · ★ 3 · AI & Automation · score 76
Install: claude install-skill aaronrau/workbench
# Kokoro G2 Transcription Loop Use the bundled runner to test the complete acoustic path: ```text Kokoro on computer → computer speaker → G2 microphones → BLE LC3 → durable capture → VAD → STT → structured Android log → score ``` ## Run the loop 1. Read [references/workbench-log-contract.md](references/workbench-log-contract.md). 2. Foreground Work Bench and use its primary connection button for a complete preflight cycle. If necessary, connect first; then tap **Disconnect**, wait for **Connect devices**, tap **Connect devices**, and wait for **Disconnect** plus fresh G2 audio frame summaries. The physical runners do this automatically and refuse to start playback if the app button or fresh post-reconnect audio is unavailable. If the connection button is initially disabled during app startup, the runner waits for the current pipeline-ready marker and retries that same visible button. Late fresh audio also satisfies the wait immediately; it is never discarded merely because connection took longer than the first probe. If initial connection work still owns the button after audio begins, the runner retries the visible **Disconnect** control until the expected disconnect marker arrives. Never infer readiness from an old connection marker. With more than one Android device attached, pass the intended Work Bench device through `--serial`. 3. Prepare the isolated computer-side runtime: ```bash python3 .agents/skills/kokoro-g2-transcr