LEHJA · build status

🎙 LEHJA — build status

Updated 2026-06-13 17:08:18 UTC · auto-refresh 60s · cron every 5 min

OVERALL (Arabic track + MT): 100%

H100 NVL: 0% GPU · active jobs 1 · VRAM 0.0 GB used / 93.1 GB free of 93.6 GB

⚙️ DATA-PREP (CPU) — speaker embeddings / speech-token extraction. GPU at 0% is normal here (ONNX runs on CPU); training starts next.

📋 Tasks — main & sub

✅ ✦. Urdu voice model (previous milestone) 100%

✅ CosyVoice2 Urdu LoRA trained + delivered · validation PASS · median CER 0.065 · model on app-server

✅ 1. Build Quranic-Arabic pipeline (canonical alignment, full tashkeel) 100%

✅ Diacritic-preserving aligner + mining/prep/train/validate scripts

✅ 2. Canonical Uthmani Quran text + alignment index 100%

✅ 6236 ayahs, full Uthmani Hafs (zabar/zer/pesh/shadda/tanwin/ghunna)

✅ Skeleton search index — rough transcript → exact ayah span

✅ 3. Mine in-domain Arabic recitation (your QTM corpus) 100%

✅ whisper-ar VAD-segment + inline Quran align · 11,900/12,000 files · 8/8 shards done

✅ clean recitation clips harvested · 11,237 clips · 1223.6 min · 3,041 distinct ayahs

✅ 4. Canonical-align → diacritized labels + confidence gate 100%

✅ perfect diacritized labels (alignment inline per clip)

✅ high-confidence gate (match≥85) + dedup · 7,238 train-grade clips

✅ 5. Train + validate + package Quranic Arabic LoRA 100%

✅ CosyVoice2 fine-tune (learn diacritized recitation) · Epoch 34/35 · loss 0.524978 · acc 0.747030

✅ validate: recites correct Arabic? (whisper-readback) · verdict: WEAK

✅ package CosyVoice2-0.5B-ar + download to app-server

✅ MT. Flagship MT benchmark (Qwen3.6-35B-A3B vs Qwen3-14B) 100%

✅ latency + Urdu/Arabic quality on real tutor sentences · results saved

📊 Arabic data harvest

Recitation clips mined: 11,237 · 1223.6 min clean audio
Train-grade (match≥85): 7,238 clips · 3,041 distinct ayahs covered
Source: your QTM Quran-recitation corpus (30,165 Arabic files available)

Approach: every clip's rough transcript is matched to the canonical Uthmani Quran text, so labels carry perfect diacritics (zabar/zer/pesh/tashdeed/tanwin/ghunna) — and non-recitation (Urdu explanation, repetition) is auto-filtered.