lemuralabs

lemura-arabic-asr-qwen3

README

License: apache-2.0

✨ Highlights

🧠 Audio-LLM architecture — audio encoder + Qwen3 decoder (generative ASR), so it uses linguistic context, not just acoustics.
🌍 Genuinely multi-dialect — trained across MSA + Gulf + Egyptian + Levantine + Iraqi + Maghrebi, including Moroccan Darija, a dialect most ASR systems struggle badly with.
🪶 Compact & efficient — 1.7B parameters, far smaller than the 7B–30B audio-LLMs, yet strongly competitive on dialectal speech.
🎯 Entity/name biasing — accepts a guiding word-list (via a system prompt) to lock in brand names, people and jargon.
🗣️ Robust to code-switching (Arabic ↔ English), common in Gulf business speech.

📊 Results — in-domain evaluation (6 standard Arabic test sets)

Average WER 22.6% across the six widely-used Arabic benchmarks, with especially strong dialectal accuracy:

Table with columns: Test set, Dialect, WER ↓
Test set	Dialect	WER ↓
MASC (clean)	mixed	10.8
CommonVoice	mixed	10.7
MGB-2	MSA / broadcast	12.9
MASC (noisy)	mixed	25.8
SADA	Gulf / Saudi	32.3
Casablanca	Maghrebi / Darija	43.2
Average	—	22.6

🌟 Strong across every dialect group, with standout MSA (MGB-2 12.9) and dialectal robustness.
🌍 Maghrebi/Darija (Casablanca 43.2) — a dialect where many strong systems score 60%+ WER. This is a real differentiator.
🪶 Achieved with a 1.7B model — punching well above its size class.

Honest note on these numbers: this is an in-domain evaluation — the model was fine-tuned on the training splits of these corpora, so the scores reflect its accuracy on the dialects and domains it has seen. On completely unseen / zero-shot audio, expect higher WER (roughly the low-30s). This is a production-oriented, dialect-broad model, not a zero-shot leaderboard submission — and it's tuned to do exactly what real deployments need: handle the dialects your users actually speak. All numbers use the official Arabic normalizer (strip punctuation/diacritics, normalize Hamza/Madda, Eastern→Western numerals) via jiwer.

📌 Model description

Task: Automatic Speech Recognition (Arabic speech → text)
Language: Arabic (ar) — MSA + Gulf, Egyptian, Levantine, Iraqi, Maghrebi/Darija
Architecture: Qwen3-ASR — audio encoder + Qwen3 LLM decoder (generative ASR), 1.7B params
Base model: Qwen/Qwen3-ASR-1.7B
Audio input: 16 kHz mono

🗂️ Training data (~1,700 hours, all 5 dialect groups)

Fine-tuned on ~1.18M clips spanning MSA and every major dialect group, including SADA (Gulf/Saudi), MASC (multi-dialect), CommonVoice-ar, MGB-2 (MSA broadcast), MoulSot & NADI (Maghrebi/Darija), and FLEURS-ar.

🧪 Training procedure

LoRA (rank 64, α 128) fine-tune of Qwen3-ASR-1.7B, bf16, then merged into standalone weights. Trained with ms-swift on NVIDIA H100s. Labels canonicalized with the standard Arabic normalizer.

🚀 How to use

With ms-swift (recommended):

bash
pip install ms-swift transformers==4.57.6 qwen-asr soundfile
swift infer \
  --model lemuralabs/lemura-arabic-asr-qwen3 \
  --val_dataset your_data.jsonl \
  --infer_backend pt --max_new_tokens 256

Dataset JSONL (one line per clip), audio 16 kHz mono:

json
{"messages": [{"role": "user", "content": "<audio>"}, {"role": "assistant", "content": ""}], "audios": ["/path/to/clip.wav"]}

🎯 Tip: entity/name biasing (recommended for best real-world accuracy)

Pass a short guiding word-list as a system message to lock in names/brands/jargon:

json
{"messages": [
  {"role": "system", "content": "الكلمات المرشدة: أبوبي، موهان، لمُرا"},
  {"role": "user", "content": "<audio>"},
  {"role": "assistant", "content": ""}],
 "audios": ["/path/to/clip.wav"]}

This alone fixes most brand/name errors in real business audio.

✅ Intended uses & limitations

Intended: transcribing Arabic speech across dialects — call centers, voice notes, media captioning, voice interfaces, especially Gulf and Maghrebi deployments.

Limitations:

Best on the dialects/domains it was trained on; fully unseen accents/domains will be harder (see honest note above).
Not a zero-shot leaderboard entry — it is optimized for broad, practical dialect coverage.
Very noisy / far-field audio degrades accuracy.
May reflect biases present in the training corpora.

🙏 Acknowledgements

Base model: Qwen/Qwen3-ASR-1.7B (Alibaba Qwen team)
Training data: SADA, MASC, CommonVoice, MGB-2, MoulSot, NADI, FLEURS

📄 License

Apache-2.0 (inherits the base model's license). Please also respect the licenses of the training datasets.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

lemuralabs

Model Tree

Base

Qwen/Qwen3-ASR-1.7B

Fine-tuned

this model

Input Modalities

Audio

Output Modalities

Text

Supported Functionality