merge-verify

Status: Current Last updated: 2026-07-17 01:36 EDT

Tier the machine-flagged utterance placements of a merged draft set against engine verdicts, rewriting promoted flags into provenance notes and exporting the rest as a review queue. Fully offline: no daemon, no models, no audio; the heavy signals arrive as a verdicts JSON produced separately.

A “merge” workflow places utterances from one source (e.g. ASR with timings) into a manually transcribed session and flags every placement it is not sure of with a %com comment. merge-verify is the second pass: given per-line verification verdicts from three engines (forced alignment confirmation, pitch banding, a machine ear), it applies a calibrated composed rule to every flagged line:

• Auto-trust: the flag is REWRITTEN into a machine-verified provenance note carrying the three signals; never silently deleted. A human transcriber note sharing the same %com tier survives verbatim ahead of the note.
• Review: the flag is left unchanged and the line is exported to review-queue.json (directly consumable as a review-campaign spot scope).
• Hold: the flag is left unchanged and the line is not queued (categories the calibration says need session-level treatment).
• Demote: a previously unflagged line whose verdicts contradict its placement gains a review flag; text and timing are never moved.

The rewritten drafts must be logically identical to the input on every main tier (a built-in preservation invariant fails the run otherwise): this pass edits only %com flags.

Quick start

batchalign3 merge-verify \
    --draft merged-drafts/ \
    --verdicts verdicts.json \
    --out verified-drafts/ \
    --flag-prefix merge

Output: one rewritten .cha per input session plus review-queue.json in --out, and a one-line tier summary on stdout.

Options

Verdicts JSON

{
 "sessions": [
  {
   "session": "S-001",
   "lines": [
    {"utterance_index": 12, "category": "other",
     "fa_mean_score": 0.52, "pitch": "child", "ear": "yes"}
   ]
  }
 ]
}

utterance_index is the 0-based ordinal over main-tier utterance lines. category is the flag taxonomy the calibration was performed over; pitch is child / adult / ambiguous; ear is yes / no. fa_mean_score orders the review queue worst-first and is never a promote/demote gate (calibration finding: disfluent child speech aligns poorly, and the aligner happily aligns the wrong voice).

The verify engines

The verdicts are produced by the three calibration-locked engines in batchalign.inference: fa_confirm (windowed MMS_FA alignment scoring), pitch_band (librosa pyin child/adult banding), and machine_ear (local audio-LLM YES/NO). Each module documents its calibration constants; changing any of them requires recalibration against blind listening verdicts.

This page last changed: 2026-07-17 (commit 03c89afe). The whole book last changed: 2026-09-16 (commit 34d249d8).