FastConformer Quran Arabic ASR — phase3_full_finetune
Fine-tuned from nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0 on tarteel-ai/everyayah.
Phase: phase3_full_finetune (Final - All layers fine-tuned)
Val WER: 0.0852
Streaming: Causal attention [128, 0] - ready for real-time inference
Progressive Unfreezing Training
- Phase 1: Top 3 encoder layers (2000 steps, LR 5e-5)
- Phase 2: Encoder layers 9-17 (3000 steps, LR 1e-4)
- Phase 3: All layers (2000 steps, LR 1e-5) ← This checkpoint
Usage
import nemo.collections.asr as nemo_asr
model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from("phase3_full_finetune_wer0.0852.nemo")
transcriptions = model.transcribe(["recitation.wav"])
print(transcriptions[0].text)
- Downloads last month
- 18