Cohere's Cohere Transcribe Arabic Beats Meta's 7B Model at 2B Parameters

Cohere releases a 2B-parameter open-source Arabic ASR model that beats Whisper Large v3 by 11 WER points and runs at 3.5x the throughput

·
·
·
Cohere's Cohere Transcribe Arabic Beats Meta's 7B Model at 2B Parameters
Read7 min
TypeNews
  • New SOTA for Arabic ASR: Cohere Transcribe Arabic tops the Open Universal Arabic ASR Leaderboard with a 25.87 average WER, beating Whisper Large v3 by 11 points.
  • Open-source, Apache 2.0: 2B-parameter model weights are free to download on Hugging Face and can run on consumer hardware or corporate servers.
  • Dialect and code-switching focus: Explicitly trained on Gulf, Egyptian, Levantine, North African, and Najdi dialects, plus Arabic-English mixed speech common in business settings.
  • 3.5x faster than Whisper: RTFx of 525 vs. 146 for Whisper Large v3, with vLLM integration for production serving via an OpenAI-compatible API endpoint.
  • Human preference: 96%: Native Arabic-speaking evaluators preferred Cohere Transcribe Arabic over Whisper in 95.8% of head-to-head tests across accuracy, dialect faithfulness, and code-switching.
  • Limitations: No automatic language detection, no timestamps or speaker diarization, and requires a VAD/noise gate to avoid hallucinating on non-speech audio.

Arabic is spoken by over 300 million people across roughly 30 recognized dialect varieties, yet it has been chronically underserved by state-of-the-art speech AI. Most existing models flatten dialectal speech into Modern Standard Arabic (MSA), the formal written register, and stumble badly on the bilingual Arabic-English mixing that is common in professional settings. Cohere is taking a direct shot at that gap with Cohere Transcribe Arabic, a new open-source ASR model that now sits at the top of the Arabic speech recognition leaderboard.

What just shipped

Cohere Transcribe Arabic is available under the Apache 2.0 license. Developers can download the weights and read quickstart implementations on Hugging Face, or access the hosted model through the Cohere API or Model Vault. The API tier includes free access subject to rate limits; managed production deployments without rate limits are available through Model Vault, priced per instance-hour with discounts for longer commitments.

The numbers that matter

Cohere Transcribe Arabic achieves the lowest average word error rate (WER) of any open-source model on the Hugging Face Arabic ASR Leaderboard, with a WER of 25.87. This is a 2.45-point improvement over the previous leader, Meta's OmniASR-LLM-7B, and an 11-point improvement over OpenAI's Whisper Large V3. To put that in perspective, OmniASR-LLM-7B is a 7-billion-parameter model; Cohere Transcribe Arabic achieves better accuracy at 2 billion parameters.

Here is how the top models compare across the six benchmark datasets:

ModelAvg WERSADACommon VoiceMASC cleanMASC noisyCasablanca
Cohere Transcribe Arabic25.8737.475.8215.5427.0749.71
OmniASR-LLM-7B28.3241.619.7519.6929.2956.46
Cohere Transcribe (prev.)30.6760.118.178.6619.0162.71
Whisper Large v336.8655.9617.8324.6634.6371.81

The gains are broad-based. Cohere Transcribe Arabic delivers the best overall WER and ranks first on four of the six composite task sets. On Casablanca, which evaluates conversational Arabic across eight dialects, it improves on OmniASR by nearly six points.

Human evaluation backed up the automated metrics. In head-to-head evaluations, Cohere Transcribe Arabic was preferred over Whisper in 95.8% of tests. Evaluators scored on three dimensions: overall accuracy, dialect faithfulness (whether the model preserved the speaker's regional variety rather than collapsing it into MSA), and robustness to code-switching.

The dialect and code-switching problem

Arabic dialect handling is a clear gap across the field. Existing models flatten dialectal speech into Modern Standard Arabic, struggle with Arabic-English code-switching, and underserve the dialects that hundreds of millions of people actually speak. This is not a niche edge case: in Gulf business settings, for example, it is entirely normal to say something like "ابغي اقدم طلب annual leave من نظام HRIS" in a single sentence, mixing Gulf Arabic with English technical terms.

Cohere's internal dialect benchmarks show the scale of the improvement:

  • On code-switched Arabic-English speech, Cohere Transcribe Arabic achieves a WER of 27.84, versus 35.37 for OmniASR-LLM-7B and 36.90 for Whisper Large v3.
  • On Gulf Arabic, it scores 24.36 WER versus 36.45 for OmniASR and 46.14 for Whisper.
  • On North African Arabic, it achieves 36.94 WER versus 47.17 for OmniASR and 71.54 for Whisper.

Under the hood

Cohere Transcribe Arabic is a 2B-parameter dedicated audio-in, text-out ASR model. It uses a conformer-based encoder-decoder architecture: a large Conformer encoder extracts acoustic representations, followed by a lightweight Transformer decoder for token generation. The Conformer (Convolution-augmented Transformer) architecture is well-suited to speech because it combines local feature extraction via convolutions with the long-range attention of Transformers.

The training data is a mix of Arabic, Arabic-accented English, and L1 English datasets, emphasizing three dimensions: dialect diversity (carefully resampled across dialect groups to prevent any single variety from dominating), Arabic-English code-switching (real bilingual speech reflecting how Arabic professionals actually speak), and acoustic variety (speech recorded across different devices, environments, and noise conditions, augmented with TTS-generated samples for underrepresented conditions).

Cohere also invested heavily in normalization. The normalizer applies full orthographic canonicalization: diacritics removal, hamza/alef variant collapsing, alif maqsura folding, Persian/Urdu letter folding, Eastern Arabic numeral transliteration, tatweel removal, punctuation normalization, and Unicode NFKC cleanup. This collapses orthographic variation that is irrelevant to transcription quality, making WER comparisons fairer across Arabic's many script conventions.

Speed that scales

Accuracy is only half the story for production systems. Cohere Transcribe Arabic achieves an RTFx (real-time factor multiple, a measure of how many times faster than real-time a model can process audio) score of 525, versus 146 for Whisper Large v3 and 66 for OmniASR-LLM-7B. That is roughly 3.5x faster than Whisper and nearly 8x faster than OmniASR, which matters enormously when processing large volumes of call-center recordings or meeting transcripts concurrently.

Cohere optimized the system around vLLM to handle high-volume speech workloads, even when audio inputs vary in length, with further updates to the runtime and model stack yielding up to 2x higher throughput. The vLLM integration means you can serve it the same way you would serve any LLM in production.

Getting started

The model is available in transformers >= 5.4.0 via the CohereAsrForConditionalGeneration class. A minimal transcription call looks like this:

makefile
from transformers import AutoProcessor, CohereAsrForConditionalGeneration
from transformers.audio_utils import load_audio
processor = AutoProcessor.from_pretrained("CohereLabs/cohere-transcribe-arabic-07-2026")
model = CohereAsrForConditionalGeneration.from_pretrained(
    "CohereLabs/cohere-transcribe-arabic-07-2026", device_map="auto"
)
audio = load_audio("your_audio.wav", sampling_rate=16000)
inputs = processor(audio, sampling_rate=16000, return_tensors="pt", language="ar")
inputs.to(model.device, dtype=model.dtype)
outputs = model.generate(**inputs, max_new_tokens=256)
text = processor.decode(outputs, skip_special_tokens=True)
print(text)

For production serving, Cohere recommends vLLM, which exposes an OpenAI-compatible /v1/audio/transcriptions endpoint. The model requires you to specify a language tag ("ar" or "en") explicitly, since it does not perform automatic language detection.

What it is and is not good at

The model's strengths are clear:

  • Multi-dialect Arabic, including Gulf, Egyptian, Levantine, Najdi, Hijazi, and North African varieties
  • Arabic-English code-switching in professional and technical contexts
  • Arabic-accented English
  • High-throughput production workloads via vLLM

There are real limitations to keep in mind:

  • The model performs best when remaining in-distribution of a single, pre-specified language. It does not feature explicit, automatic language detection and exhibits inconsistent performance on code-switched audio.
  • The model does not support output timestamps or speaker diarization.
  • Like most AED (audio encoder-decoder) speech models, Cohere Transcribe Arabic is eager to transcribe even non-speech sounds, and benefits from prepending a noise gate or VAD (voice activity detection) model to prevent floor noise from turning into hallucinations.

Why this matters beyond the benchmark

The real significance here is not just a leaderboard position. Arabic has been a second-class citizen in speech AI despite being the fifth most spoken language in the world. Most Arabic ASR work has focused on MSA, which is the formal written standard but is not how most people actually talk day-to-day. Saudi Arabia alone is home to three major dialect groups and many more linguistic subgroups. A model that genuinely handles this diversity, runs on-premise under a permissive license, and does so at production-grade throughput is a qualitatively different kind of tool.

The practical use cases are broad: call-center analytics for Gulf telecoms, meeting transcription for MENA enterprise software, voice interfaces for Arabic-language applications, and compliance recording in regulated industries. Cohere Transcribe Arabic runs efficiently on consumer hardware, with no reliance on external APIs or cloud services, and is available under a permissive license. For organizations in the region with data sovereignty requirements, that last point is often the deciding factor.

Trending
  • No trending articles

Comments

avatar

Next Reads