Audar-ASR-V1-Flash is the edge tier of Audar's Arabic-first speech-recognition family —
the same in-house Arabic training program as Audar-ASR-V1-Turbo,
delivered in a fast ~0.6B-decoder model for real-time captioning and on-device use. It recasts
transcription as audio-conditioned next-token prediction (a language-model decoder, not CTC/transducer),
and is built on a permissively-licensed open-weight audio-LLM foundation, then adapted in-house through Audar's Arabic training program — the contribution is the adaptation, not the foundation:
🧱 Large-scale bilingual pretraining — 300,000+ hours of labeled audio, primarily Arabic and English (MSA + Gulf, Egyptian,
Levantine, Maghrebi; code-switching; diverse channels).
🎯 Dialect-targeted fine-tuning with hardness and multi-task sampling.
🧠 KTO preference alignment (Kahneman-Tversky Optimization) from trained native-Arabic annotators.
It transcribes MSA and every major Arabic dialect, code-switched Arabic–English, and English, across
30 languages, and runs on CPU / GPU / edge via 🤗 Transformers or GGUF. For maximum accuracy on
the hardest dialectal audio, use the larger Turbo tier.
Built on a permissively-licensed open-weight audio-LLM foundation; the adaptation, data, and
alignment are Audar's. Full method and results: Audar-ASR-V1 Technical Report. Runs via Transformers, llama.cpp / GGUF, and vLLM.
Model summary
Model
Audar-ASR-V1-Flash — Arabic-first generative ASR (edge tier)
Task
Automatic speech recognition (audio → text)
Approach
Generative ASR — audio encoder + language-model decoder
Training
built on an open-weight audio-LLM foundation; adapted via a curriculum — 300k+ hrs bilingual pretraining → dialect-targeted SFT → KTO alignment
Decoder parameters
596,049,920 (0.60B)
Audio encoder parameters
186,376,192 (0.19B)
Total parameters
782,426,112 (0.78B, bf16)
Audio input
16 kHz mono; 30 s context (longer audio is chunked/streamed)
Languages
Arabic (MSA + Gulf/Egyptian/Levantine/Maghrebi dialects) + English + 28 more
Open Universal Arabic ASR Leaderboard — full standings
Flash is evaluated end-to-end on all six leaderboard test sets (full test splits, not sampled), with
the leaderboard-equivalent normalizer — the same harness and protocol as every other row (Audar rows below are the leaderboard
maintainers' independent reproduction, Aug 2026 normalization). Audar-ASR-V1-Flash scores 32.0 avg WER
at just 0.78B parameters — on par with Qwen3-ASR-1.7B (2× its size) and ahead of
Voxtral-Small-24B, Whisper-large-v3, and every CTC baseline.
Audar's accuracy tier, Turbo, is #1.
Per-dataset WER % across all six sets, plus the two composite averages. Lower is better; Avg WER
is the ranking metric. Flash and Turbo (Ours) in bold; bold cell = best in column.
#
Model
Avg WER
Avg CER
SADA
CV-18
MASC-clean
MASC-noisy
MGB-2
Casablanca
1
Audar-ASR-V1-Turbo (Ours, 2.35B)
23.17
9.20
28.92
8.09
16.73
27.19
11.08
47.02
2
CohereLabs/cohere-transcribe-arabic-07-2026
25.87
11.80
37.47
5.82
19.60
27.07
15.54
49.71
3
omnilingual-asr/omniASR_LLM_7B
28.32
12.52
41.61
8.75
19.69
29.29
14.13
56.46
4
omnilingual-asr/omniASR_LLM_3B
29.96
13.77
46.18
9.15
19.90
30.03
14.22
60.27
5
omnilingual-asr/omniASR_LLM_1B
29.96
13.40
43.84
9.55
20.03
30.26
15.34
60.68
6
CohereLabs/cohere-transcribe-03-2026
30.67
16.37
60.11
8.17
8.66
19.01
25.33
62.71
7
Qwen/Qwen3-Omni-30B-A3B-Instruct
30.71
13.67
44.82
11.46
21.47
30.85
13.09
62.55
8
Audar-ASR-V1-Flash (Ours, 0.78B)
32.04
13.43
44.36
15.38
22.56
34.19
17.05
58.71
9
nvidia-conformer-ctc-large-arabic (lm)
32.91
13.84
44.52
8.80
23.74
34.29
17.20
68.90
10
omnilingual-asr/omniASR_LLM_300M
32.96
14.84
51.38
12.03
20.66
32.45
16.58
64.64
11
google/gemma-4-E4B-it
32.98
13.71
43.40
19.65
24.86
33.59
17.72
58.63
12
Qwen/Qwen3-ASR-1.7B
33.36
12.33
45.53
16.90
24.37
34.29
16.57
64.47
13
mistralai/Voxtral-Small-24B-2507
34.47
15.29
50.82
15.25
23.96
34.43
16.03
66.30
14
nvidia-conformer-ctc-large-arabic (greedy)
34.74
13.37
47.26
10.60
24.12
35.64
19.69
71.13
15
google/gemma-4-E2B-it
35.87
15.34
46.23
23.76
27.47
36.15
20.72
60.87
16
openai/whisper-large-v3
36.86
17.21
55.96
17.83
24.66
34.63
16.26
71.81
17
omnilingual-asr/omniASR_CTC_3B
37.78
19.79
69.85
14.19
21.48
34.60
18.96
67.58
18
omnilingual-asr/omniASR_CTC_7B
38.12
20.91
72.69
12.47
21.08
35.04
20.43
67.02
19
facebook/seamless-m4t-v2-large
38.16
17.03
62.52
21.70
25.04
33.24
20.23
66.25
20
omnilingual-asr/omniASR_CTC_1B
39.29
20.47
71.42
17.55
22.76
35.73
19.96
68.32
21
openai/whisper-large-v3-turbo
40.05
18.87
60.36
25.73
25.51
37.16
17.75
73.79
22
openai/whisper-large-v2
40.20
19.55
57.46
21.77
27.25
38.55
25.17
71.01
23
Qwen/Qwen3-ASR-0.6B
42.19
16.23
53.75
28.28
31.34
42.63
25.45
71.68
24
openai/whisper-large
42.57
20.49
63.24
26.04
28.89
40.79
24.28
72.18
25
mistralai/Voxtral-Mini-3B-2507
42.58
19.90
63.65
22.12
28.37
41.27
22.56
77.52
26
asafaya/hubert-large-arabic-transcribe
45.50
17.35
67.82
8.01
32.94
50.16
37.51
76.53
27
openai/whisper-medium
45.57
22.27
67.71
28.07
29.99
42.91
29.32
75.44
28
nvidia-Parakeet-ctc-1.1b-concat
46.54
23.88
70.70
26.34
30.49
45.95
24.94
80.80
29
omnilingual-asr/omniASR_CTC_300M
46.65
21.86
78.11
27.90
28.40
43.26
26.85
75.35
30
nvidia-Parakeet-ctc-1.1b-universal
51.96
25.19
73.58
40.01
36.16
50.03
30.68
81.30
31
microsoft/VibeVoice-ASR
52.99
28.95
69.83
44.25
32.95
52.43
25.10
93.37
32
facebook/mms-1b-all
54.54
21.45
77.48
26.52
38.82
57.33
39.16
87.95
33
openai/whisper-small
55.13
21.68
78.02
24.18
35.93
56.36
48.64
87.64
34
whitefox123/w2v-bert-2.0-arabic-4
58.13
27.62
87.34
41.79
37.82
53.28
40.66
87.88
35
jonatasgrosman/wav2vec2-large-xlsr-53-arabic
60.98
25.61
86.82
23.00
42.75
64.27
56.29
92.72
36
speechbrain/asr-wav2vec2-commonvoice-14-ar
65.74
30.93
88.54
29.17
49.10
69.57
64.37
93.68
Flash — per-dataset detail (full test sets)
Both metrics, for the six leaderboard sets and the composite average.
Dataset
WER %
CER %
SADA
44.36
23.57
CommonVoice-18
15.38
4.91
MASC-clean
22.56
7.48
MASC-noisy
34.19
12.52
MGB-2
17.05
7.97
Casablanca
58.71
24.14
Average (6-set)
32.04
13.43
Use Flash for real-time and on-device transcription; step up to
Turbo when you need the lowest
error on heavy dialectal or long-form audio — Turbo is #1 on the leaderboard (23.2 % avg WER) and
cuts Flash's average WER by ~8.9 pp, with the biggest gains on SADA (44.4→28.9) and MGB-2 (17.1→11.1).
Our leaderboard numbers were produced with the qwen-asr package, which implements this model's I/O protocol natively — and were independently reproduced by the leaderboard maintainers with this exact code:
language="Arabic" makes the package prefill language Arabic<asr_text> into the prompt, so the model never free-runs language identification.
The model's no-speech verdict (language None<asr_text>) is mapped to an empty transcript; without this, non-speech audio (music, silence) can produce repetition loops.
max_new_tokens=256 and bf16 are the exact decode settings behind our published numbers.
If you use raw transformers (below), you must strip the language <Lang><asr_text> output prefix yourself and expect degraded scores on non-speech-heavy data.
🤗 Transformers inference
Ships self-contained modeling code, so trust_remote_code=True is required.
python
1# pip install "transformers>=4.57" torch librosa2import re, torch, librosa
3from transformers import AutoProcessor, AutoModelForCausalLM
45repo ="audarai/Audar-ASR-V1-Flash"6proc = AutoProcessor.from_pretrained(repo, trust_remote_code=True)7model = AutoModelForCausalLM.from_pretrained(8 repo, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map="cuda:0",9).eval()1011SYSTEM ="فرّغ الكلام العربي التالي."# "Transcribe the following Arabic speech."12audio, _ = librosa.load("clip.wav", sr=16000, mono=True)1314conv =[15{"role":"system","content": SYSTEM},16{"role":"user","content":[{"type":"audio"}]},# audio placeholder (a list, not "<audio>")17]18text = proc.apply_chat_template(conv, tokenize=False, add_generation_prompt=True)19inputs = proc(text=text, audio=audio, sampling_rate=16000, return_tensors="pt").to(model.device)20inputs["input_features"]= inputs["input_features"].to(model.dtype)# features are fp32 → cast to bf162122out = model.generate(**inputs, max_new_tokens=440, do_sample=False)23hyp = proc.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0]24print(re.sub(r"^\s*language\s+[A-Za-z]+\s*(?:<asr_text>)?\s*","", hyp).strip())
Language steering: the Arabic auto-dialect prompt above needs no dialect hint. For other
languages use e.g. "Transcribe the following speech.".
Long audio (>30 s): split at ~30 s boundaries (see the streaming section).
⚡ vLLM inference (GPU serving)
Flash also serves on vLLM with an OpenAI-compatible API. vLLM
implements the Qwen3-ASR architecture natively, so it serves the repo's bf16 model.safetensors
directly — no conversion, lossless (FLEURS AR/EN CER ≈ 3.3 %).
1. Install (audio support required)
The stock vLLM image ships no audio codecs; add PyAV + librosa + soundfile:
dockerfile
1FROM vllm/vllm-openai:v0.24.02RUN pip install --no-cache-dir av librosa soundfile
vLLM auto-detects the model; ~1.6 GB weights fit on any ≥ 8 GB GPU.
3. Transcribe
Send 16 kHz mono audio as base64 input_audio with the Arabic system prompt, decode greedily
(temperature: 0). POST /v1/audio/transcriptions (multipart) or /v1/chat/completions both work:
Output note: Flash prefixes raw output with a language <Lang><asr_text> tag — strip it
client-side (same as the Transformers example above):
re.sub(r"^\s*language\s+[A-Za-z]+\s*(?:<asr_text>)?\s*", "", text).strip().
💻 GGUF inference (llama.cpp)
Audar-ASR runs on llama.cpp via the multimodal (mtmd) path: a quantized decoder GGUF plus a
BF16 audio projector (mmproj). Build a recent llama.cpp (with Qwen3-ASR support), then:
⚠️ The audio projector (mmproj) must stay BF16 — the encoder's ClippableLinear is numerically
sensitive, so F16/Q8 measurably degrade quality. The decoder quantizes normally.
GGUF variants
File
Approx. size
Notes
Audar-ASR-V1-Flash-Q4_K_M.gguf
~0.40 GB
Smallest; best for edge/offline
Audar-ASR-V1-Flash-Q8_0.gguf
~0.64 GB
Near-lossless, CPU-friendly (recommended)
Audar-ASR-V1-Flash.gguf (BF16)
~1.20 GB
Full precision decoder
mmproj-Audar-ASR-V1-Flash.gguf
~0.38 GB
BF16 audio encoder — required, keep BF16
Prefer a managed endpoint? The Audar-ASR family is also available via the
Audar API/SDK — streaming, speaker-attributed transcription, and
diarization, production-hosted.
🎙️ Real-time streaming
The 30 s-context model streams via LocalAgreement-2: as audio arrives, the trailing window is
re-decoded each hop and a word is committed only once two consecutive decodes agree on it — giving
stable, low-latency incremental output on both the Transformers and GGUF paths. Audar's production
realtime engine serves the same policy over an OpenAI-Realtime-compatible WebSocket with model-based
endpointing.
🌍 Languages, dialects & tasks
Primary: Arabic — MSA and dialectal (Gulf/Emirati, Egyptian, Levantine, Maghrebi), plus
code-switched Arabic–English; dialect-faithful orthography from audio alone.
Also: English + 28 additional languages.
Task: transcription (audio → UTF-8 text), prompt-steerable for language/formatting.
Intended use & limitations
Intended use. Live captioning and subtitles, voice assistants/agents, meeting and call-center
transcription, media/broadcast, accessibility — cloud, on-prem, or offline/edge.
Limitations.
Maghrebi / Moroccan Darija (Casablanca) is the hardest condition for all systems.
Heavily code-switched telephony and low-SNR audio degrade accuracy relative to clean MSA.
Long recordings can drift; chunk at sentence boundaries for best results.
Not evaluated for, and must not be used for, covert speaker identification.
📜 License
Released under the AudarAI Open License v1.0 — commercial use, redistribution, and
fine-tuning/quantization permitted; ship the license and keep notices. See
audarai.com/license/audarai-open-license-v1.0.
Citation
bibtex
1@misc{audar-asr-flash-2026,
2 title = {Audar-ASR-V1: A Multilingual, Arabic-First Generative Speech Recognition Foundation Model},
3 author = {AudarAI},
4 year = {2026},
5 note = {Audar-ASR-V1-Flash},
6 url = {https://github.com/AudarAI/Audar-ASR-V1/blob/main/report/Audar-ASR-V1-Technical-Report.pdf}
7}
About AudarAI
Leading Arabic-First Multilingual Audio Intelligence
AudarAI starts with Arabic — and expands to the world.
We are building advanced multilingual audio intelligence that helps individuals, enterprises, and
governments communicate across languages, cultures, and borders. By combining Arabic-first speech
technology with global multilingual AI, AudarAI transforms voice into understanding, interaction,
and connection.
Our work spans speech recognition, speech understanding, voice-enabled digital assistants,
human-computer interaction, and intelligent audio systems designed for real-world impact. From
empowering people to access technology in their native language to helping organizations
communicate globally, AudarAI is shaping a future where every voice can be heard, understood, and
connected.
Arabic-first. Multilingual by design. Human-centered at heart.
-f59e0b){kind=icon}
