OpenTranscribe — Cloud ASR, Speakers & Voiceprints

OpenTranscribe — Cloud ASR, Speakers & Voiceprints A workflow diagram generated by Archify. 01 / API / plan 02 / Cloud vendors (audio leaves) 03 / This server One cloud-asr task Voiceprint + match Upload · user's audio/video · API / plan Upload user's audio/video Pipeline plan · ASR config + detector · API / plan › One cloud-asr task · refuses no timing Pipeline plan ASR config + detector refuses no timing Cloud ASR · 9 vendors, one factory · Cloud vendors (audio leaves) › One cloud-asr task Cloud ASR 9 vendors, one factory Cloud diarizer · pyannote.ai or adapter · Cloud vendors (audio leaves) › One cloud-asr task · parallel_cloud Cloud diarizer pyannote.ai or adapter parallel_cloud Merge · assign_speakers · This server › One cloud-asr task · word / segment Merge assign_speakers word / segment Voiceprint · cpu embedding stage · This server › Voiceprint + match Voiceprint cpu embedding stage Match · same as local ASR · This server › Voiceprint + match Match same as local ASR same audio text (+ vendor speakers) turns only Legend Agent logic Context / trace Cloud service External system

Who detects the speakers

  • • inline_asr: the transcription vendor's own labels (tasks/transcription/cloud_asr.py)
  • • parallel_cloud: pyannote.ai, or another saved provider through services/diarization/asr_adapter.py (its text is discarded, it bills a full transcription)
  • • Automatic + a text-only model: this server after transcription on the full image; refused on lite

Timing decides the merge

  • • word: each word gets the overlapping speaker; segment: each segment does and its words follow
  • • none (gpt-4o-transcribe): refused with asr_no_timing_for_speakers; there is no forced alignment
  • • A cloud plan never falls back to local Whisper; a failed cloud diarizer keeps the text

Voiceprints stay local

  • • services/voiceprint_extractor.py on the cpu queue (speaker_embedding_task), on full and lite alike
  • • Same matching and clustering downstream as local ASR; see the Speaker Clustering diagram