---
library_name: nemo
license: cc-by-4.0
language:
- ar
- arz
- apc
- ary
- afb
pipeline_tag: automatic-speech-recognition
metrics:
- wer
tags:
- automatic-speech-recognition
- speech-recognition
- speech-to-text
- stt
- transcription
- speech
- audio
- asr
- arabic
- arabic-asr
- arabic-speech
- dialectal-arabic
- msa
- modern-standard-arabic
- gulf-arabic
- egyptian-arabic
- levantine-arabic
- maghrebi-arabic
- multi-dialect
- nemo
- nemo-toolkit
- nvidia
- fastconformer
- fastconformer-ctc
- conformer
- ctc
- efficient
- lightweight
- compact
- on-device
- edge
- cpu-inference
- real-time
- low-latency
- low-resource
model-index:
- name: lemura-arabic-asr-lite
results:
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
name: Open Universal Arabic ASR Leaderboard (average of 6 sets)
type: open-universal-arabic-asr
metrics:
- type: wer
value: 25.08
name: Average WER
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
name: MASC (clean)
type: masc-clean
metrics:
- type: wer
value: 7.27
name: WER
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
name: Common Voice 18 (Arabic)
type: common-voice-18-ar
metrics:
- type: wer
value: 9.74
name: WER
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
name: MGB-2
type: mgb-2
metrics:
- type: wer
value: 14.33
name: WER
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
name: MASC (noisy)
type: masc-noisy
metrics:
- type: wer
value: 23.65
name: WER
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
name: SADA
type: sada
metrics:
- type: wer
value: 37.28
name: WER
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
name: Casablanca
type: casablanca
metrics:
- type: wer
value: 58.24
name: WER
---

# lemura-arabic-asr-lite
   
**A compact, dialect-aware Arabic speech-recognition model — leaderboard-tier accuracy at ~18× fewer parameters.**
    
[ Overview](#-overview) · [ Model Details](#-model-details) · [ Benchmarks](#-benchmarks) · [ Efficiency](#-efficiency) · [ Usage](#-usage) · [ Live Demo](https://huggingface.co/spaces/lemuralabs/lemura-arabic-asr-demo)
---
## Overview
**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 and fine-tuned on **\~2,900 hours** of Arabic spanning **MSA** and the **Gulf, Egyptian, Levantine, and Maghrebi** dialect groups.
**Highlights**
- **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 Details
| | |
|---|---|
| **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](https://github.com/Natural-Language-Processing-Elm/open_universal_arabic_asr_leaderboard) (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 |
## 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
```python
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](https://huggingface.co/spaces/lemuralabs/lemura-arabic-asr-demo)**
## Credits
- Acoustic foundation: **NVIDIA FastConformer**
- Evaluation: the **[Open Universal Arabic ASR Leaderboard](https://github.com/Natural-Language-Processing-Elm/open_universal_arabic_asr_leaderboard)** official code