Sabeq-ASR

| | | | | | | | | | | |

Sabeq-ASR is an Arabic automatic speech-recognition model built on a hybrid FastConformer architecture with both RNN-T (Transducer) and CTC decoders, fine-tuned with NVIDIA NeMo.

It achieves a state-of-the-art results with an average WER of 20.40% and CER of 9.12% across the six test sets of the Open Universal Arabic ASR Leaderboard using RNN-T beam search, and 21.11% WER / 9.79% CER with greedy decoding.

🧭 Model overview

• Architecture: FastConformer encoder with hybrid RNN-T + CTC decoding heads.
• Base model: FastConformer (113M parameters).
• Tokenizer: Unigram, vocabulary size 1024.
• Audio input: 16 kHz mono.
• Decoding: greedy or RNN-T beam search; beam search uses beam_size=20 and score normalization.
• Framework: NVIDIA NeMo.

Efficiency

Metric	Value
Average RTF	0.01631
Model memory	0.63 GB
Peak inference memory	0.45 GB

An RTF of 0.016 means the model transcribes audio roughly 61× faster than real time (1 / 0.01631), making it suitable for batch transcription and low-latency applications.

✨ Highlights

State-of-the-art Arabic transcription accuracy. Across the six public test sets of the Open Universal Arabic ASR Leaderboard, Sabeq-ASR reaches an average WER of 20.40% and CER of 9.12% using RNN-T beam search surpassing Jais-ASR, Sygma-ASR-Arabic, Audar-ASR-V1-Turbo, and Cohere Transcribe Arabic with both greedy and beam decoding. It leads on SADA22, MASC-clean, and MASC-noisy.

Designed around dialectal and noisy Arabic. The model combines a FastConformer encoder with dual RNN-T and CTC decoding heads, and was fine-tuned on speech covering Modern Standard Arabic alongside Gulf, Egyptian, Levantine, and Maghrebi dialects, including challenging real-world acoustic conditions.

Small, fast, and flexible. At only 113M parameters, Sabeq-ASR is roughly 50× smaller than Jais-ASR (6B), 20× smaller than Audar-ASR-V1-Turbo (2.35B), and 16× smaller than Sygma-ASR-Arabic and Cohere Transcribe Arabic (2B) — yet it still achieves the lowest average WER on the leaderboard. It delivers this accuracy without the overhead of a generative language-model decoder, and beam search (beam_size=20) yields a further 0.71-point WER gain over greedy decoding.

Simple to deploy. The model loads directly through NVIDIA NeMo with a few lines of Python, works with standard 16 kHz mono audio, and is distributed under the permissive CC-BY-4.0 license for both research and commercial applications.

---

📊 Benchmarks

Results on the six public test sets used by the Open Universal Arabic ASR Leaderboard. Lower is better. Our results were produced with the leaderboard's eval script.

Model	Avg WER	Avg CER	SADA22	CV-18	MASC-clean	MASC-noisy	MGB-2	Casablanca
Sabeq-ASR (Beam)	20.40	9.12	28.29	7.01	5.43	18.13	9.89	53.63
Sabeq-ASR (Greedy)	21.11	9.79	29.74	7.36	5.79	18.75	10.27	54.77
IFM/Jais-ASR	22.82	8.31	30.52	8.59	16.14	25.47	9.60	46.61
abedbanna/Sygma-ASR-Arabic	23.01	9.52	33.63	5.24	15.92	25.10	13.48	44.67
audarai/Audar-ASR-V1-Turbo	23.17	9.23	28.92	8.09	16.73	27.19	11.08	47.02
CohereLabs/cohere-transcribe-arabic-07-2026	24.19	11.46	36.83	5.27	16.35	25.65	15.33	45.68

Bold = best in column. Sabeq-ASR obtains the lowest average WER and is the best model on SADA22, MASC-clean, and MASC-noisy.

Greedy vs. beam decoding

Dataset	Greedy WER	Greedy CER	Beam WER	Beam CER
MASC-clean	5.79	2.62	5.43	2.35
MASC-noisy	18.75	7.53	18.13	7.03
SADA22	29.74	14.10	28.29	12.90
MGB-2	10.27	4.31	9.89	4.04
CommonVoice 18 (Arabic)	7.36	2.26	7.01	2.09
Casablanca	54.77	27.91	53.63	26.33
Average	21.11	9.79	20.40	9.12

RNN-T beam search improves average WER by 0.71 percentage points and average CER by 0.67 percentage points over greedy decoding.

💻 Usage

pip install "nemo_toolkit[asr]"

Greedy decoding

from huggingface_hub import hf_hub_download
import nemo.collections.asr as nemo_asr

model_path = hf_hub_download(
    repo_id="MostafaAhmed98/Sabeq-ASR",
    filename="model.nemo",
)
# RNNT used as default
model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from(model_path)
output = model.transcribe(['sample.wav'])
print(output[0].text)

RNN-T beam search

This is the decoding configuration used for the reported 20.40% average WER.

from huggingface_hub import hf_hub_download
import nemo.collections.asr as nemo_asr

model_path = hf_hub_download(
    repo_id="MostafaAhmed98/Sabeq-ASR",
    filename="model.nemo",
)

model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from(model_path)

decoding_cfg = model.cfg.decoding

with open_dict(decoding_cfg):
    decoding_cfg.strategy = "beam"
    decoding_cfg.beam.beam_size = 20
    decoding_cfg.beam.score_norm = True
    decoding_cfg.beam.return_best_hypothesis = True

model.change_decoding_strategy(decoding_cfg, decoder_type="rnnt")

transcripts = model.transcribe(["/content/Algeria_11213_1.wav"])
print(transcripts[0].text)

📜 License

This model is released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license.

Citation

@misc{sabeq-asr-2026,
  title  = {Sabeq-ASR: A Hybrid FastConformer Model for Arabic Speech Recognition},
  author = {Mostafa Ahmed},
  email  = {mostafa.ahmed00976@gmail.com},
  year   = {2026},
  url    = {MostafaAhmed98/Sabeq-ASR}
}