lemura-arabic-asr

A compact, dialect-aware Arabic speech-recognition model — leaderboard-tier accuracy at ~18× fewer parameters.

🧭 Overview · 📊 Benchmarks · ⚡ Efficiency · 💻 Usage · 🎙️ Live Demo

---

🧭 What it is

lemura-arabic-asr is a compact multi-dialect Arabic ASR model built for accuracy and efficiency. It is a FastConformer-CTC acoustic model (115M parameters) adapted in-house from the NVIDIA FastConformer foundation, fine-tuned on **2,900 hours** of Arabic spanning MSA and the Gulf, Egyptian, Levantine, and Maghrebi dialect groups.

• 🪶 Small & fast — ~115M parameters. Runs comfortably on CPU and in real time, no GPU required.
• 🗣️ Dialect-aware — trained across five Arabic dialect groups, not MSA-only.
• 🎧 Robust on real audio — strongest on broadcast, conversational, and Gulf/MSA speech.
• 🔓 Open & simple — a single .nemo file, loadable in a few lines with NVIDIA NeMo.

Model summary

Model	lemura-arabic-asr — compact multi-dialect Arabic ASR
Task	Automatic speech recognition (audio → text)
Architecture	FastConformer encoder + CTC decoder
Parameters	~115M
Foundation	adapted from NVIDIA FastConformer; ~2,900 h Arabic fine-tuning
Audio input	16 kHz mono (auto-resampled)
Languages	Arabic — MSA + Gulf / Egyptian / Levantine / Maghrebi
Runtime	NVIDIA NeMo — CPU · GPU · real-time
License	CC-BY-4.0

📊 Benchmarks

Evaluated on all six Open Universal Arabic ASR Leaderboard test sets using the official leaderboard code (same normalizer, same WER metric).

Per-set WER (%)

SADA	Common Voice 18	MASC-clean	MASC-noisy	MGB-2	Casablanca	Average
37.28	9.74	7.27	23.65	14.33	58.24	25.08

In context (Average WER %, lower is better)

Model	Params	Avg WER
cohere-transcribe-arabic-07-2026	~2.0B	25.87
lemura-arabic-asr	~0.12B	25.08*
omniASR_LLM_7B	7B	28.32
Qwen3-Omni-30B-A3B	30B	30.71
nvidia-conformer-ctc-large-arabic (lm)	0.6B	32.91
Qwen3-ASR-1.7B	1.7B	33.36

* Transparency note. lemura-arabic-asr was fine-tuned on training splits that overlap several of these test corpora (SADA, MASC, CV, MGB-2). The leaderboard is designed to measure zero-shot generalization, so this 25.08 is best read as an in-domain result, not a strict zero-shot ranking — a fair zero-shot number is a few points higher. Its unqualified strengths are efficiency and Gulf / MSA / broadcast robustness; the frontier for every system remains Maghrebi Darija (Casablanca).

⚡ Why it matters: efficiency

The top leaderboard systems are large generative audio-LLMs (2–30B parameters) that need GPUs. lemura-arabic-asr reaches a comparable accuracy tier with a ~115M-parameter CTC model:

	lemura-arabic-asr	Typical top systems
Parameters	~115M	2B – 30B
Hardware	CPU or GPU	GPU
Latency	real-time	seconds / clip
Footprint	~0.4 GB	4 – 60 GB

That makes it practical for on-device, low-cost, and high-throughput Arabic transcription where the big models are impractical.

💻 Usage

import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.restore_from("asr_final.nemo")
print(model.transcribe(["audio.wav"]))   # 16 kHz mono

Try it live — no install: 🎙️ lemura-arabic-asr demo

🙏 Credits

• Acoustic foundation: NVIDIA FastConformer.
• Evaluation: the Open Universal Arabic ASR Leaderboard official code.