Real-time transcription: how to recover state when the WebSocket drops mid-sentence
tresch_24 · reddit · 2026-09-17
A developer building a real-time transcription pipeline (audio → backend WebSocket → Smallest AI Pulse → transcript events → app state) raises a core engineering question: when the socket dies mid-sentence and reconnects 800ms later, should you replay audio, start fresh, or wait for a late final from the old stream? Replaying can duplicate transcript text; not replaying may eat the exact word you needed — worse when a partial already mutated UI/state. The author's proposed fix: give each segment its own seq/id plus idempotent downstream handling, so one reconnect doesn't turn into state archaeology. He'd take another 300ms of latency over duplicated state any day.
More from coding & agent
- Matt Pocock proposes /pr skill: every model creates garbage PR bodies today — mattpocockuk · 2026-09-17
- Codex Agents Pile Up 76GB of Debug Files, Burning Through 5x Quota Overnight — craigbalding · 2026-09-17
- Build the agent setup first, pick the model second: a Linux + Tailscale + llama.cpp stack guide — max_paperclips · 2026-09-17
- Dev claims LLM memory is basically solved using LLM plus Jev for recall — JoshuaJBouw · 2026-09-17
- AI agents can now control Google Home as MCP server opens to third-party assistants — emmanuelvivier · 2026-09-17
- Bindu Reddy Teases Upgraded Agent Integrating Jev, Launching Tomorrow — bindureddy · 2026-09-17