This is a byte-exact mirror of NVIDIAs Arabic FastConformer hybrid Transducer-CTC model with Punctuation, Capitalization, and Diacritics output (Tashkeel).

Why a mirror

This mirror exists for availability redundancy in case the upstream repo is renamed, gated, or removed. It is byte-identical to the upstream checkpoint and carries no modifications.

Source

Upstream repo: nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0
Pinned upstream commit SHA: 7f32349d952f42a28dce979ba73270aa2bbdfa89
Mirrored on: 2026-05-06
Mirrored by: dev-ahmedhany
License: CC-BY-4.0 (unchanged)

File integrity

sha256sum stt_ar_fastconformer_hybrid_large_pcd_v1.0.nemo
d29d19d7c054a5fc010ac6815e9cbb0dd1b21a30e0a7f7f2982e1fecaf0c3e31  stt_ar_fastconformer_hybrid_large_pcd_v1.0.nemo

Verify after download:

bash

huggingface-cli download dev-ahmedhany/stt_ar_fastconformer_hybrid_large_pcd_v1.0-mirror \
  stt_ar_fastconformer_hybrid_large_pcd_v1.0.nemo --local-dir .
sha256sum stt_ar_fastconformer_hybrid_large_pcd_v1.0.nemo
# expected: d29d19d7c054a5fc010ac6815e9cbb0dd1b21a30e0a7f7f2982e1fecaf0c3e31

Attribution (CC-BY-4.0 requirement)

Original model and weights by NVIDIA. All credit for the model architecture, training, and weights belongs to NVIDIA. This mirror simply preserves bit-for-bit access to the upstream artifact.

If you use this model, please cite the original NeMo / FastConformer papers as listed on the upstream model card.

Usage

Identical to upstream:

python

import nemo.collections.asr as nemo_asr
m = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.from_pretrained(
    "dev-ahmedhany/stt_ar_fastconformer_hybrid_large_pcd_v1.0-mirror"
)
print(m.transcribe(["audio.wav"])[0].text)

For complete usage details, training data, evaluation WER per dataset, and limitations, see the upstream model card linked above.

License

This mirror is distributed under CC-BY-4.0, the same license as the upstream NVIDIA repository. You may use, redistribute, and modify under the terms of CC-BY-4.0 with proper attribution to NVIDIA.