Streaming Transcription Service
Design a reliable service that turns live or uploaded audio into ordered, editable transcripts with speaker attribution.
Problem statement
For this exercise, assume users start a live transcription session or upload a completed recording. Live sessions should expose revisable partial text and then a stable final transcript with speaker labels.
Users may correct finalized text without destroying the machine-produced revision, export the latest approved revision, and delete a session. Design ingestion, segmentation, recognition, ordering, revision, storage, and recovery.
The service may call an existing speech-recognition model; training that model and building audio-capture hardware are outside scope.
What to build
- Clients can stream numbered audio chunks or upload a recording and safely resume interrupted ingestion.
- Clients can observe low-latency partial text and retrieve one ordered finalized transcript with speaker labels.
- Authorized users can correct finalized spans, preserve revision history, and export the latest approved transcript.
- Authorized users can request deletion of audio and transcript artifacts and observe completion status.
At scale
- Support 100,000 concurrent live sessions averaging 32 KiB of audio per second each.
- Publish partial text within 1.5 seconds at p95 and finalize 99% of sessions within two minutes after audio ends.
- Do not lose acknowledged audio chunks; make duplicate chunks harmless and preserve transcript order across reconnects.
- Retain source audio for 30 days and transcript revisions for one year unless a verified deletion request shortens both.