Review the user's dictations

The user dictates with Blurt all day. Their takes are the best test data there is: real voice, real pauses, real vocabulary. This skill turns them into fixes.

Prerequisites

• Settings → 听写 → 「保留录音」 (record.dictation.keep_audio) on. Then each take keeps its audio ($BLURT_HOME/dictation/audio/<id>.wav) and its streamed pieces (cuts, pauses, every polish window's input and output) until the history retention (30 days). Without it only the text is there: say so and review the text alone.
• Never write ~/.blurt except through dictation_lab.py (cases) — no hand edits of the history or config.
• The audio is the user's voice. Allowed off this Mac (user, 2026-10-28): disputed clips only, to OpenAI's speech models through OpenRouter, at most $10 in total — cloud_ref.py next to this file enforces the cap (ledger spend.json, clips ≤ 120 s). Nothing else leaves the Mac.

Cloud second opinion (disputed clips)

OPENROUTER_API_KEY=$(grep '^OPENROUTER_API_KEY=' /Users/john/workspace/ordo_projects/ordo_ai/ordo-backend/.env.prod | head -1 | cut -d= -f2- | tr -d "\"'") \
  python3 -I .claude/skills/blurt-dictation-review/cloud_ref.py CLIP.wav [--language zh]

Default models: openai/whisper-large-v3 (≈ $0.03 / audio hour) and openai/gpt-transcribe (≈ $0.27 / hour). Never print the key. On Chinese, Whisper is weaker than the local Qwen3-ASR (rare names 12 % vs 4.5 % errors), so read it as a vote: two cloud models agreeing against Blurt is strong evidence of a recognition error; one alone is not. Cut the disputed span out of a long take (ffmpeg -ss/-to) rather than sending it all.

Steps

1. uv run skills/blurt/scripts/dictation_lab.py review --since <ISO date> (or --last N). Each kept take comes back with typed, raw (streamed ASR), heard_once (the whole take re-heard in one pass by the installed model, no cuts), heard_diff, the pieces with pause_after and each window's polish_in / text, and flags: fell_back:<reason>, heard_differently, seam_marks, spelled_letters, numeral_format. For a take without audio, read ~/.blurt/dictation/history.jsonl (raw / polished / text / fell_back / reason).
2. Read every take, flagged or not, as a careful human editor would: what did the user most likely say? Classify each problem:
   • heard wrong (ASR): homophone (买点/埋点), term (空中云会/空中云汇), dropped or invented words. If heard_once has it right, the streaming (cuts, prior) caused it; if not, the model.
   • pause punctuation: a mark where the user only paused to think (look at pause_after ≥ 0.6 s and the window that wrote it); or missing marks in dense speech.
   • formatting: 三D → 3D, 四十八小时 → 48 小时, C L I → CLI, V 二 → V2, spaces.
   • polish: fell back (why — the guard's reason), dropped / changed / translated words, copied the text before the cursor.
   • user: they misspoke — not a bug. When unsure whether the user or the model is wrong, re-run the take through other configurations (dictation_lab.py from-history ID then rerun CASE --config asr=qwen3-asr-1.7b-mlx,mode=whole …) and the cloud second opinion below.
3. Save each real error as a case with its ground truth: dictation_lab.py from-history ID --corrected "<what they meant>" --tags term,punctuation --note "<why>". Tags: term, code_switch, punctuation, filler, dropped, made_up, segmentation, clipped, slow (+ format, homophone).
4. Report to the user in a short table: take, what was typed, what was meant, class, cause. Then group by cause and propose fixes — model choices, prompt, guard, seams, vocabulary — never word lists or regex rules for language (docs/design/dictation.md §Rules vs model).
5. When fixing: measure before / after on all saved cases (dictation_lab.py run --config mode=windows --json) plus the eval sets (polish.py eval, tests/fixtures/polish_overdelete.jsonl), and the real-take replay of the history text. Ship only what improves the cases without regressions; say the numbers.