cjm-transcription-core

A frontend-agnostic core for the audio transcription workflow — composes isolated capability workers (audio conversion, VAD segmentation, batch transcription, persistence) into a headless pipeline, with a CLI as its first driver.

Modules

• cjm_transcription_core.boundaries — Wall-clock-aware segment boundary computation: group VAD speech chunks into segments cut at silence-gap midpoints. Pure logic — no capability calls. Final home of the algorithm originally validated in cjm-transcription-audio-segment's AudioSegmentService.compute_segment_boundaries (that library is retired to cj-mills_deferred/).
• cjm_transcription_core.cli — The CLI driver — the workflow core's first (and currently only) frontend.
• cjm_transcription_core.curation — Collection curation vocabulary (hub v0, e5849229): the journaled update/delete
• cjm_transcription_core.emission — Graph-root emission (CR-18 revolution 2): a completed source EMITS Source -> AudioSegment -> Transcript into the shared context graph — the graph BEGINS at transcription (where-graph-begins resolution: ingestion is the first EXTENDER that plants the root). Deterministic identity tuples make emission idempotent: re-runs (cache hits included) collide into verified no-ops instead of duplicating roots (the E13 hazard, relocated into graph creation and discharged).
• cjm_transcription_core.models — Data shapes for the transcription pipeline: run configuration + the run-manifest result containers. The run manifest is the pipeline's durable output record: which sources were processed, how they were segmented, and where each segment's transcription landed (capability data DBs remain the authoritative text store; the manifest records the run's shape + provenance pointers). It is a deliberate proto-bundle — the CR-20 provenance-bundle infrastructure is expected to absorb/replace it.
• cjm_transcription_core.pipeline — The headless transcription pipeline: VAD analysis -> boundary computation -> segment cutting -> per-segment model-input conversion -> transcription, composed over capability workers via the substrate's JobQueue. Between-stage outputs are threaded manually (run job -> read result -> submit next); the per-segment fan-out rides a CR-16 ports Composition with OutputRef bindings (this module was the real-world consumer of the original submit_sequence piping gap — pass-2 evidence in claude-docs/pass-2-evidence.md). HITL approval seams use the cheapest viable form (log + optional CLI prompt) per the cores-cluster guard-rails; each seam carries its 5-field HITL-assist annotation in its docstring.

API

cjm_transcription_core.boundaries

• compute_segment_boundaries function — Group VAD chunks into segments cut at silence-gap midpoints.

cjm_transcription_core.cli

• build_parser function — Build the CLI parser (subcommands: run).
• expand_sources function — Expand CLI source arguments into the ordered media-file list for a run.
• expand_sources_with_collections function — Expand CLI sources AND keep the folder-source gesture as collection
• load_capabilities function — Discover manifests + load each requested capability.
• main function — CLI entry point (console script: cjm-transcription-core).
• parse_max_concurrent function — Parse repeatable --max-concurrent NAME=N values into a per-capability cap map.
• parse_transcriber_spec function — Parse one --transcriber spec into a (capability, MODEL)-instance load directive.
• run_command function — Execute the run subcommand: full pipeline over the given audio files.

cjm_transcription_core.curation

• apply_curation function — Replay one collection-curation op: deletes -> updates -> wires.
• collection_members function — A collection's member Sources (PART_OF edges; unordered by design —
• collection_order function — Walk the materialized order, when one exists (typed EdgeQuery reads —
• confirm_collection function — Discharge a proposed collection's flag (ae3464fc: the explicit human
• curation_replay_handlers function — The curation verb's replay registration (unioned into
• file_sources function — File existing Sources into a collection (create-or-attach; the hub's
• journal_curation function — Apply one curation act and journal it as a collection-curation op.
• list_collections function — Enumerate the graph's Collection nodes (the hub's grouping corpus).
• refile_members function — Move members between collections (the Supernova carve-out: select
• rename_collection function — Rename a collection — which IS merge when the new title already exists.
• set_collection_order function — Materialize (or repair) a collection's order — the curation op ae3464fc

cjm_transcription_core.emission

• build_collection_emission function — Build the Collection layer payload for one declaration (pure; no
• build_source_emission function — Build the graph-root payload for one source (pure; no capability calls).
• emit_collections_graph function — Idempotently emit the run's collection declarations (verb
• emit_source_graph function — Idempotently emit one source's graph root through the task channel.
• transcription_replay_handlers function — The transcription core's replay vocabulary (DEC 426658f1, replay stays DOMAIN-OWNED).

cjm_transcription_core.models

• CollectionDecl class — A collection declaration riding a run (ae3464fc: the folder-source
• PipelineConfig class — Configuration for one transcription pipeline run.
• RunManifest class — Durable record of one pipeline run (proto-bundle; see CR-20).
• SegmentRecord class — One segment of a source audio file, with per-transcriber transcripts.
• SourceResult class — Pipeline result for one source audio file.
• new_run_id function — Generate a unique, sortable run id.

cjm_transcription_core.pipeline

• acquire_speaker_turns function — Diarize the full source and persist the source-keyed turns artifact.
• analyze_vad function — Run VAD analysis on one model-ready audio file (task channel: vad/detect_speech).
• build_segment_composition function — Build the per-source fan-out composition: N independent [preprocess→]convert→(T× transcribe) pipes.
• collect_capability_info function — Record capability identity + data-DB pointers for the run manifest (provenance).
• confirm_seam function — HITL approval seam in its cheapest viable form (log + optional CLI prompt).
• convert_for_vad function — Convert a source to MODEL-READY audio for VAD via the ffmpeg convert action.
• cut_segments function — Cut the source audio at the computed boundaries via ffmpeg segment_audio.
• normalize_vad_result function — Normalize a typed VAD result into sorted speech chunks + the reported duration.
• probe_duration function — Probe a media file's duration via the ffmpeg capability's get_info action.
• records_from_composition function — Fold a completed segment composition back into SegmentRecords.
• run_pipeline function — Run the transcription pipeline over the given sources, in order.
• run_source function — Run the full pipeline for one source: VAD → boundaries → cut → [preprocess →] convert → transcribe.
• submit_and_wait function — Submit one capability job, wait for it, and return its result (raise on failure).
• tier1_segment_checks function — Tier-1 deterministic pre-filters for the boundary-review seam (no AI).
• tier1_transcript_checks function — Tier-1 deterministic pre-filters for the transcript-review seam (no AI).

Dependencies

Depends on: cjm-capability-primitives, cjm-context-graph-layer, cjm-context-graph-primitives, cjm-substrate, cjm-transcript-graph-schema, cjm-transcription-adapter-interface Used by: cjm-transcription-tui, cjm-workflow-hub-tui