OpenTranscribe — Post-Transcription Enrichment

OpenTranscribe — Post-Transcription Enrichment A workflow diagram generated by Archify. 01 / Core 02 / Speaker Identity 03 / Search 04 / LLM Enrichment Dispatched on finalize Speaker identity chain Finalize · transcript written to Postgres · Core Finalize transcript written to Postgres Voiceprint · embeddings, local or cloud · Speaker Identity › Dispatched on finalize · gpu or cpu queue Voiceprint embeddings, local or cloud gpu or cpu queue Gender/Age · speaker attribute detection · Speaker Identity › Speaker identity chain Gender/Age speaker attribute detection Speaker ID · LLM suggestion, manual approve · Speaker Identity › Speaker identity chain · never auto-applied Speaker ID LLM suggestion, manual approve never auto-applied Reclustering · selective, on request · Speaker Identity › Speaker identity chain Reclustering selective, on request Indexing · OpenSearch whole-doc + chunks · Search › Dispatched on finalize Indexing OpenSearch whole-doc + chunks Summarization · LLM summary · LLM Enrichment › Dispatched on finalize Summarization LLM summary Topics · LLM topic extraction · LLM Enrichment › Speaker identity chain Topics LLM topic extraction Analytics · talk-time, stats · LLM Enrichment › Speaker identity chain Analytics talk-time, stats context for LLM on request Legend Agent logic Context / trace External system

Voiceprinting already covers cloud ASR

  • • Local ASR: embeddings come inline from PyAnnote's native centroids (no extra GPU pass)
  • • Cloud ASR (no native embeddings): a dedicated extract_speaker_embeddings_task runs on the CPU queue — same voiceprint/fingerprint outcome, different path

LLM speaker ID is a suggestion, never a write

  • • Gender/age detection runs first so the LLM has that context
  • • Confidence-scored suggestions surface for manual approval only — repo-wide rule

Independent fan-out

  • • Indexing, summarization, and analytics all dispatch right after finalize
  • • None of them block the speaker-identity chain or each other