Audio Pipeline

  period: project
  converted: 2026-05-28

Overview

3-type audio classification and processing pipeline; work audio never goes to Sandpit; Phase 7 gated on Otter MCP scout.

3-type audio classification: work=extract+discard (never Sandpit), music=retain in Sandpit, general=retain in Sandpit.

Why: Work audio is transit — value is in the synthesis (daily digest), not the file. Music audio is source material for the analysis pipeline and must be retained.

How to apply: When routing audio, always check type first. Work .m4a stays in inbox/ until transcribed, then discarded. Music .m4a moves to Sandpit. Specs: the-government/information_reference/runtime-state/reference_audio-pipeline-spec.md + the-government/information_reference/reference_work-audio-protocol.md (originally at colonel/saitama-foundation/kakashi-work/work-audio-protocol.md, now dissolved into the-government/).

Current State (2026-05-15)

• Work audio protocol written (Otter-style daily digest, team vocabulary, vocal profiles)

• Phase 7 (Whisper transcription) not started — gated on Otter MCP scout

• Phase 9 (vocal profile embeddings) backlogged — gated on Phase 7

• 18 work .m4a files in inbox/pending-review awaiting Phase 7

• Career - Zoopla Interview-.m4a in inbox/ — needs debrief from Michael

• Sandpit routing correction logged: prior spec incorrectly sent work audio to Sandpit

Phases

  PhaseDescriptionStatusGate

    7Whisper transcription + WhisperX diarisationNot startedOtter MCP scout verdict
    9Speaker vocal profiles (pyannote + ECAPA-TDNN)BackloggedPhase 7 stable

Trending Monitor Config

Asset file: projects/2026-05-audio-pipeline/trending-monitor-config.json — raw config data embedded for reference. Original JSON preserved at source path.

◆ hinata · projects/audio-pipeline.html · phase-18 flatten