🔥 RE-TRANSCRIBING — Arabic recitation → exact canonical Quran text with full tashkeel.

✅ ✦. LEHJA voice models (previous — COMPLETE) 100%

✅ Urdu TTS (PASS, CER 0.065) + Quranic Arabic v2 (owner-validated 95%)

✅ MT bench: Qwen3.6-35B-A3B wins

✅ 1. Setup & research (QTM-Canada) 100%

✅ Mapped stack (.NET + Postgres), content & audio model

✅ Safety backup of 47,058 transcripts saved

✅ 26.5k→37.1k ar+en+ur chunks exported; all audio staged on H100

🔄 2. Pass A — Re-transcribe + Quran-align (Arabic→exact tashkeel) 15%

✅ Pipeline built: whisper (forced-lang, VAD) + canonical Quran alignment

✅ Verification samples — anchor span-fill, 86% Arabic→exact diacritized text

🔄 Full run (37,119 chunks) · 5,557/37,119 done · 1,305 Quran-matched

⬜ 3. Pass B — Qwen3.6-35B-A3B strict cleanup (non-Quran en/ur/ar) 0%

⬜ fix-only-obvious-errors, never change meaning · 0/~16,430

🔄 4. Pass C — Speaker diarization (flag >2 speakers + runaway recordings) 3%

🔄 count speakers/chunk → new table agent_audio_speaker_stat · 1,278/37,119

⬜ 5. DB writeback (in-place, old backed up) 0%

⬜ agent_audio_chunk.transcript updated · 0/37,119

⬜ speaker-stat table populated · 0 rows

Audio: 37,119 chunks / ~631 hours (Arabic 24,056 · English 10,556 · Urdu 2,507)
Method: whisper re-transcribe + canonical Quran alignment (Arabic→exact diacritized) + Qwen strict cleanup (conversational) + speaker diarization.
Safety: old transcripts backed up; updates in place; runaway recordings (up to 18h) capped at 15 min + flagged.