LEHJA training

🎙 LEHJA — Urdu voice-model training (live)

Updated: 2026-06-12 21:50:01 UTC · cron har 5 min · page 60s auto-refresh

OVERALL PIPELINE: 55%

GPU: 0% busy · jobs: 1 · VRAM: 0.0 GB used / 93.1 GB free / 93.6 GB total

📋 Tasks (main + sub)

1. H100 Setup ✅ 100%

✅ SSH access + box verification — complete (2 min lagi)

✅ Python deps + CosyVoice repo install — complete (6 min lagi)

✅ CosyVoice2-0.5B model download (5.3 GB) — complete (5 min lagi)

✅ Whisper CUDA-12 libs fix (silent-fail pakra gaya tha) — complete (14 min lagi)

2. Data Acquisition ✅ 100%

✅ Catalogued Urdu files VM101→H100 (2,667 files) — complete (3 min lagi)

✅ Full pool transfer (93,166 recordings, 17 GB) — complete (9 min lagi)

✅ Catalog language-map (45,185 entries) — complete (1 min lagi)

✅ Public FLEURS Urdu+English — 3,637 clips — complete (26 min lagi)

⊘ CommonVoice Urdu (HF-gated, zaroorat nahi) — skipped

3. In-domain Urdu Mining (asal awaazein) ✅ 100%

✅ Known-Urdu transcription — 2,667 files, 6 GPU workers — complete (3 min lagi) · 2,668/2,668 files · 17,605 clips · 1340.2 min saaf Urdu

✅ Language-scan — 53,054 unknown files, 16 GPU workers — complete (18 min lagi) · 53,054/53,054 scanned · naye Urdu-candidates: 3,219

✅ Round-2: naye mile files ki transcription — complete (14 min lagi) · 3,219/3,219

4. Training-Data Prep ⬜ 0%

⬜ Manifests merge + dedup + train/dev split — pending

⬜ wav.scp / text / utt2spk / spk2utt — pending

⬜ Speaker embeddings (campplus) — pending

⬜ Speech tokens extraction — pending

⬜ Parquet packaging — pending

5. CosyVoice2 URDU TRAINING ⬜ 0%

⬜ LLM module fine-tune (Urdu phonetics) — pending · ~10 epochs planned

⬜ Checkpoint averaging (best-5) — pending

⬜ Inference smoke-test (Urdu bolta hai?) — pending

6. Validation ⬜ 0%

⬜ Urdu test-set synthesis (50 jumlay) — pending

⬜ Whisper-readback intelligibility gate — pending

⬜ Zero-shot voice-clone spot test (apki awaaz) — pending

7. Delivery ⬜ 0%

⬜ Sab artifacts app-server par download (model+data+logs) — pending

⬜ COMPLETION EMAIL → info@ifcondition.com → AAP H100 BAND KAR DEN — pending

📊 Harvest

Saaf Urdu (training-ready): 1340.2 min in-domain (17,605 clips) + 3,637 public clips
Scan se naye Urdu-candidate files: 3,219 / 53,054 scanned · worker failures: 0

Kaam jari hai — mukammal hone par email: info@ifcondition.com