cjm-transcription-core 0.0.1

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.

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. Ships in-package as the cjm-transcription-core console script so the driver can never skew from the core. GUI presentation drivers come later and consume the same pipeline module; they never reimplement it (CLI-first / headless-core principle).
• 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.
• tests_manual.measure_cold_parallel_overlap_e2e — Cold-run parallel-overlap measurement (stage-5 closeout; the carried G11
• tests_manual.validate_oom_backstop_e2e — OOM backstop stress test (stage-3 ledger): Voxtral-Small-24B on a 24GB GPU.
• tests_manual.validate_stage8_sn1_both_transcribers_e2e — Stage-8 SN-I BOTH-TRANSCRIBER parity — the voxtral Option-C migration (build step 2).
• tests_manual.validate_stage8_sn1_e2e — Stage-8 SN-I closeout — the whisper Option-C migration validated at scale.

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).
• load_capabilities function — Discover manifests + load each requested capability (default instance).
• 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.
• run_command function — Execute the run subcommand: full pipeline over the given audio files.

cjm_transcription_core.emission

• build_source_emission function — Build the graph-root payload for one source (pure; no capability calls).
• emit_source_graph function — Idempotently emit one source's graph root through the task channel.

cjm_transcription_core.models

• 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

• 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).

tests_manual.measure_cold_parallel_overlap_e2e

• lane_intervals function — (start, end) UTC intervals for completed jobs of one instance.
• main function
• merge function — Merge overlapping intervals so a side can't double-count itself.
• overlap_seconds function — Total seconds where any interval of A overlaps any interval of B.
• self_overlap_max function — Max simultaneously in-flight jobs within one lane + seconds at depth >= 2.

tests_manual.validate_oom_backstop_e2e

• empirical_rows function
• main function
• pipe_composition function

tests_manual.validate_stage8_sn1_both_transcribers_e2e

• check_discovery function — Milestone 1: {whisper, voxtral} compatible set, manifest-surface-based.
• journal_max_seq function
• journal_rows function
• main function
• run function

tests_manual.validate_stage8_sn1_e2e

• journal_max_seq function
• journal_rows function
• main function
• run function

Dependencies

Depends on: cjm-substrate