mohammed/fastconformer-quran-ar

FastConformer Quran Arabic ASR — phase1_top3

Fine-tuned from nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0 on tarteel-ai/everyayah.

Phase: phase1_top3 | Val WER: 0.0038

The .nemo file is saved with full bilateral context so transcribe() works out of the box. For real-time streaming, apply causal attention at inference time (see streaming script).

Offline usage

import nemo.collections.asr as nemo_asr
model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from("model.nemo")
result = model.transcribe(["recitation.wav"])
print(result[0].text)

Streaming usage

Load with restore_from(), then:

model.change_attention_model("rel_pos_local_attn", att_context_size=[128, 0])
# Then use cache_aware_stream_step() for frame-by-frame inference