lemura-arabic-asr-qwen3

Lemura Labs' Arabic-first generative speech-recognition model — dialect-broad, production-oriented.

Built for the dialects your users actually speak.

Overview · Benchmarks · Transformers · ms-swift · Entity Biasing · Training · Sibling model

---

What it is

lemura-arabic-asr-qwen3 is an Arabic-first generative speech-recognition model. It recasts transcription as audio-conditioned next-token prediction — an audio encoder feeding a Qwen3 language-model decoder, rather than a CTC or transducer objective — so it resolves ambiguity with linguistic context, not acoustics alone.

Fine-tuned from Qwen/Qwen3-ASR-1.7B, it transcribes Modern Standard Arabic and five major dialect groups — Gulf/Khaleeji, Egyptian, Levantine, Iraqi and Maghrebi/Darija — in a single compact 1.7B model.

• Audio-LLM architecture — generative ASR; context disambiguates where acoustics cannot.
• Genuinely multi-dialect — ~1,700 hours across MSA plus five dialect groups, including Moroccan Darija, which most ASR systems handle badly.
• Dialect-targeted fine-tuning — LoRA adaptation focused on dialectal orthography, proper nouns and code-switching, merged back into standalone weights.
• Entity and name biasing — accepts a guiding word-list to lock in brands, people and jargon.
• Robust to code-switching — Arabic ↔ English, common in Gulf business speech.

Built for the hard part of Arabic ASR: real dialectal speech, code-switching, and named entities.

Built on a permissively-licensed open-weight audio-LLM foundation; the adaptation, data curriculum and dialect coverage are Lemura Labs'.

Model summary

Model	lemura-arabic-asr-qwen3 — Arabic-first generative ASR
Task	Automatic speech recognition (audio → text)
Approach	Generative ASR — audio encoder + Qwen3 language-model decoder (audio-conditioned next-token prediction)
Base model	Qwen/Qwen3-ASR-1.7B
Training	LoRA (rank 64, α 128) dialect fine-tune in bf16, merged to standalone weights — ~1.18M clips / ~1,700 h across 5 dialect groups
Architecture class	Qwen3ASRForConditionalGeneration (model_type: qwen3_asr)
Audio encoder	24 layers · d_model 1024 · 16 heads · 128 mel bins
Total parameters	1.7B (bf16)
Audio input	16 kHz mono; 30 s context window (longer audio is chunked)
Languages	Arabic — MSA + Gulf / Egyptian / Levantine / Iraqi / Maghrebi, with Arabic–English code-switching
Runtime	transformers · ms-swift — GPU
License	Apache-2.0

Benchmarks

Open Universal Arabic ASR Leaderboard — full standings

Per-dataset WER % across all six leaderboard test sets, zero-shot. Lower is better; Avg WER is the ranking metric. Our sibling model lemura-arabic-asr-lite is evaluated here with the official leaderboard code; this model's base, Qwen/Qwen3-ASR-1.7B, is row 12. These are zero-shot numbers — the in-domain table above is a different measurement and the two must not be compared directly.

#	Model	Params	Avg WER	SADA	CV-18	MASC-clean	MASC-noisy	MGB-2	Casablanca
1	lemuralabs/lemura-arabic-asr-lite (Ours)	0.12B	25.08	37.28	9.74	7.27	23.65	14.33	58.24
2	CohereLabs/cohere-transcribe-arabic-07-2026	~2B	25.87	37.47	5.82	19.60	27.07	15.54	49.71
3	omnilingual-asr/omniASR_LLM_7B	7B	28.32	41.61	8.75	19.69	29.29	14.13	56.46
4	omnilingual-asr/omniASR_LLM_3B	3B	29.96	46.18	9.15	19.90	30.03	14.22	60.27
5	omnilingual-asr/omniASR_LLM_1B	1B	29.96	43.84	9.55	20.03	30.26	15.34	60.68
6	CohereLabs/cohere-transcribe-03-2026	~2B	30.67	60.11	8.17	8.66	19.01	25.33	62.71
7	Qwen/Qwen3-Omni-30B-A3B-Instruct	30B	30.71	44.82	11.46	21.47	30.85	13.09	62.55
8	nvidia-conformer-ctc-large-arabic (lm)	0.6B	32.91	44.52	8.80	23.74	34.29	17.20	68.90
9	omnilingual-asr/omniASR_LLM_300M	0.3B	32.96	51.38	12.03	20.66	32.45	16.58	64.64
10	google/gemma-4-E4B-it	4B	32.98	43.40	19.65	24.86	33.59	17.72	58.63
11	Qwen/Qwen3-ASR-1.7B (our base, zero-shot)	1.7B	33.36	45.53	16.90	24.37	34.29	16.57	64.47
12	mistralai/Voxtral-Small-24B-2507	24B	34.47	50.82	15.25	23.96	34.43	16.03	66.30
13	nvidia-conformer-ctc-large-arabic (greedy)	0.6B	34.74	47.26	10.60	24.12	35.64	19.69	71.13
14	google/gemma-4-E2B-it	2B	35.87	46.23	23.76	27.47	36.15	20.72	60.87
15	openai/whisper-large-v3	1.5B	36.86	55.96	17.83	24.66	34.63	16.26	71.81
16	omnilingual-asr/omniASR_CTC_3B	3B	37.78	69.85	14.19	21.48	34.60	18.96	67.58
17	omnilingual-asr/omniASR_CTC_7B	7B	38.12	72.69	12.47	21.08	35.04	20.43	67.02
18	facebook/seamless-m4t-v2-large	2.3B	38.16	62.52	21.70	25.04	33.24	20.23	66.25
19	omnilingual-asr/omniASR_CTC_1B	1B	39.29	71.42	17.55	22.76	35.73	19.96	68.32
20	openai/whisper-large-v3-turbo	0.8B	40.05	60.36	25.73	25.51	37.16	17.75	73.79
21	openai/whisper-large-v2	1.5B	40.20	57.46	21.77	27.25	38.55	25.17	71.01
22	Qwen/Qwen3-ASR-0.6B	0.6B	42.19	53.75	28.28	31.34	42.63	25.45	71.68
23	openai/whisper-large	1.5B	42.57	63.24	26.04	28.89	40.79	24.28	72.18
24	mistralai/Voxtral-Mini-3B-2507	3B	42.58	63.65	22.12	28.37	41.27	22.56	77.52
25	asafaya/hubert-large-arabic-transcribe	0.3B	45.50	67.82	8.01	32.94	50.16	37.51	76.53
26	openai/whisper-medium	0.8B	45.57	67.71	28.07	29.99	42.91	29.32	75.44
27	nvidia-Parakeet-ctc-1.1b-concat	1.1B	46.54	70.70	26.34	30.49	45.95	24.94	80.80
28	omnilingual-asr/omniASR_CTC_300M	0.3B	46.65	78.11	27.90	28.40	43.26	26.85	75.35
29	nvidia-Parakeet-ctc-1.1b-universal	1.1B	51.96	73.58	40.01	36.16	50.03	30.68	81.30
30	microsoft/VibeVoice-ASR	—	52.99	69.83	44.25	32.95	52.43	25.10	93.37
31	facebook/mms-1b-all	1B	54.54	77.48	26.52	38.82	57.33	39.16	87.95
32	openai/whisper-small	0.24B	55.13	78.02	24.18	35.93	56.36	48.64	87.64
33	whitefox123/w2v-bert-2.0-arabic-4	0.6B	58.13	87.34	41.79	37.82	53.28	40.66	87.88
34	jonatasgrosman/wav2vec2-large-xlsr-53-arabic	0.3B	60.98	86.82	23.00	42.75	64.27	56.29	92.72
35	speechbrain/asr-wav2vec2-commonvoice-14-ar	0.1B	65.74	88.54	29.17	49.10	69.57	64.37	93.68

Competitor rows are the published Open Universal Arabic ASR Leaderboard standings, reproduced for context; they are not our measurements. Casablanca (Moroccan Darija) is the hardest set for every system.

Transformers inference

The full-precision bf16 weights are published at the repo root (safetensors). Qwen3-ASR is supported natively in recent Transformers — no trust_remote_code needed.

# pip install "transformers>=4.57.6" torch librosa
import torch, librosa
from transformers import AutoProcessor, Qwen3ASRForConditionalGeneration

repo = "lemuralabs/lemura-arabic-asr-qwen3"
processor = AutoProcessor.from_pretrained(repo)
model = Qwen3ASRForConditionalGeneration.from_pretrained(
    repo, dtype=torch.bfloat16, device_map="auto"
).eval()

audio, _ = librosa.load("clip.wav", sr=16_000)          # 16 kHz mono

messages = [{"role": "user", "content": [{"type": "audio", "audio": audio}]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=text, audio=audio, sampling_rate=16_000, return_tensors="pt").to(model.device)

out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

ms-swift inference

For batch transcription over a dataset, ms-swift is the fastest path:

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:

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

Entity and name biasing

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

{"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 and name errors in real business audio.

Notes

• Long audio: the audio encoder uses a 30 s window (chunk_length: 30); chunk longer inputs client-side and concatenate the transcripts.
• Sample rate: input must be 16 kHz mono; resample before inference.
• Decoding: greedy (do_sample=False) is recommended for transcription; sampling adds hallucination risk.
• Precision: weights are bf16; fp16 works on older GPUs, but bf16 is what the model was trained and merged in.

Training data

~1,700 hours / ~1.18M clips spanning MSA and every major dialect group:

Corpus	Contribution
SADA	Gulf / Saudi dialectal speech
MASC	Multi-dialect, clean and noisy
CommonVoice-ar	Read speech, mixed dialects
MGB-2	MSA broadcast
MoulSot, NADI	Maghrebi / Moroccan Darija
FLEURS-ar	Read speech, MSA

Training procedure

LoRA (rank 64, α 128) fine-tune of Qwen3-ASR-1.7B in bf16, then merged into standalone weights. Trained with ms-swift on NVIDIA H100s. Labels canonicalized with the standard Arabic normalizer (punctuation and diacritics stripped, Hamza/Madda normalized, Eastern→Western numerals).

Languages, dialects and tasks

• Primary: Arabic — MSA and dialectal (Gulf/Khaleeji, Egyptian, Levantine, Iraqi, Maghrebi/Darija), plus code-switched Arabic–English; emits dialect-faithful orthography from audio alone.
• Task: transcription (audio → UTF-8 text), optionally steered by a guiding word-list for entities.
• Audio: 16 kHz mono, 30 s window.

Intended use and limitations

Intended use. Transcribing Arabic speech across dialects — contact centres, voice notes, media captioning, voice agents and accessibility, especially Gulf and Maghrebi deployments; cloud or on-prem.

Limitations.

• Best on the dialects and domains it was trained on; fully unseen accents or domains will be harder (see the in-domain note above).
• Maghrebi / Moroccan Darija remains the hardest condition, as it is for every system.
• Not a zero-shot leaderboard entry — optimized for broad, practical dialect coverage.
• Very noisy or far-field audio degrades accuracy; long recordings must be chunked.
• Not evaluated for, and must not be used for, covert speaker identification.
• May reflect biases present in the training corpora.

Related model

lemura-arabic-asr-lite — ~115M-parameter FastConformer-CTC, CPU and real-time capable, #2 of 36 zero-shot on the Open Universal Arabic ASR Leaderboard.

	-qwen3 (this model)	-lite
Parameters	1.7B	~115M
Approach	Generative audio-LLM	FastConformer-CTC
Hardware	GPU	CPU or GPU
Strength	In-domain dialect depth, entity biasing	Zero-shot generality, edge deployment
Reported WER	22.6 (in-domain)	25.08 (zero-shot, leaderboard)

License

Apache-2.0, inheriting the base model's license. Please also respect the licenses of the training datasets.

Citation

@misc{lemura_arabic_asr_qwen3_2026,
  title  = {lemura-arabic-asr-qwen3: Generative Multi-Dialect Arabic Speech Recognition},
  author = {Lemura Labs},
  year   = {2026},
  url    = {https://huggingface.co/lemuralabs/lemura-arabic-asr-qwen3}
}

Acknowledgements

• Base model: Qwen/Qwen3-ASR-1.7B (Alibaba Qwen team)
• Training data: SADA, MASC, CommonVoice, MGB-2, MoulSot, NADI, FLEURS
• Evaluation: the Open Universal Arabic ASR Leaderboard official code

---

About Lemura Labs

Arabic-first, efficiency-first speech intelligence.

Lemura Labs starts with the dialects people actually speak.

We build compact, deployable speech and language models — accuracy at a size and cost that works outside the datacentre. Our work spans speech recognition, dialectal language modelling, and efficient inference for real-world deployment, from edge devices to on-prem clusters.

Dialect-first. Efficient by design. Built to deploy.

Hugging Face · developer@aivf.io

© 2026 Lemura Labs · Licensed under Apache-2.0