Three problems, three trade-offs
Voice updates sound easy until you start. Three problems jumped out:
- Latency. Engineers won't wait 8 seconds for transcript.
- Accuracy. Code identifiers (
useEffect,gRPC) trip every off-the-shelf model. - Storage. A 90-second WAV is 1.5MB. At 2,400 teams, daily, that's a lot of S3.
What we did
- Streaming Whisper instead of batch — first words show up in 200ms.
- Glossary prompt loaded per workspace — caught the code identifiers.
- Re-encode to Opus + 30-day expiry — cut storage 9×.
What's next
Speaker diarisation. Right now Loop assumes one voice per update. For pair updates, we want both names attributed automatically. Estimated ship: Q3.
RC
by Renzo Castellanos
Staff Engineer, Loop
