metadata
title: Super-fast Arabic ASR – 93.3 % accuracy in 250 epochs
emoji: 🎙️
colorFrom: blue
colorTo: green
sdk: docker
pinned: false
license: apache-2.0
tags:
- automatic-speech-recognition
- arabic
- asr
- nemo
- finetuned
- pytorch
- audio
- transformers
🔊 Arabic FastConformer Hybrid – Finetuned for Gulf & MSA
One-liner inference & ready-to-use Docker API
</div>
📌 Model Card
| Base model | NVIDIA stt_ar_fastconformer_hybrid_large_pc |
| Fine-tuning data | 2 000+ hours synthetic + real Gulf & MSA speech |
| Vocab | 1 024 BPE sub-words (Arabic + English digits) |
| WER (eval set) | 6.7 % |
| CER (eval set) | 2.9 % |
| Accuracy | 93.3 % |
| Sample rate | 16 kHz mono |
🚀 5-Second Inference
from nemo.collections.asr.models import EncDecHybridRNNTCTCBPEModel
model = EncDecHybridRNNTCTCBPEModel.restore_from("alaatiger989/Arabic_Finetuned_ASR_Nemo")
transcript = model.transcribe(["my_audio.wav"])[0][0]
print(transcript) # -> "محمد أحمد عبد الرحمن"