Cohere Releases Transcribe Arabic, an Open-Source Model for Arabic Speech Recognition
Cohere's new open-source Transcribe Arabic model reports the lowest error rate of any open-source Arabic speech-recognition model, ahead of Whisper and OmniASR.
What Cohere released
Cohere released Cohere Transcribe Arabic as an open-source model, built on its 2B frontier Automatic Speech Recognition (ASR) model that the company says it released earlier this year. Cohere says the new model is designed for the realities of Arabic in business and developer settings: dialect variation, bilingual Arabic-English speech, code-switching, and domain-specific vocabulary. Cohere Transcribe Arabic is available under the Apache 2.0 license; developers can download the weights and quickstart implementations from Hugging Face, or access the hosted model through the Cohere API or Model Vault.
Why Arabic is hard for speech models
Cohere's post frames Arabic as a language whose diversity complicates transcription: more than 300 million people speak Arabic as a mother tongue across roughly 30 recognized varieties, and Saudi Arabia alone contains three major dialect groups. Modern Standard Arabic (MSA) provides a common written standard, the company notes, but everyday speech varies significantly across dialects, and code-switching, using Arabic and non-Arabic vocabulary in the same conversation, is common in professional settings. Cohere says this combination has left Arabic under-served relative to English. The company built Cohere Transcribe Arabic by taking Cohere Transcribe, launched in March, and training it extensively on data spanning Arabic dialects, professional language, code-switching, and varied acoustic conditions.
What Cohere reports measuring
Cohere reports that Transcribe Arabic achieves the lowest average word error rate (WER) of any open-source model on the Hugging Face Arabic ASR Leaderboard, at 25.87, 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. On the Open Universal Arabic ASR leaderboard, which evaluates zero-shot multi-dialect generalization across six test sets spanning MSA, Egyptian, Gulf, Levantine, and Maghrebi dialects, Cohere says its model ranks first on four of the six composite task sets; on Casablanca, which tests conversational Arabic across eight dialects, it improves on OmniASR by nearly six points, and on the crowd-sourced Common Voice set covering 25 dialects, it reduces word error rate by more than two points. In human evaluations judging overall accuracy, dialect faithfulness, and code-switching robustness, Cohere reports native Arabic speakers preferred the model over Whisper in 95.8% of head-to-head tests, while the company's own headline figure states reviewers preferred it to Whisper in 96% of tests overall. Cohere also reports the model improved substantially over Cohere Transcribe on English spoken with an Arabic accent, preferred in 77.2% of tests, and was roughly comparable to Whisper on those inputs, preferred in 52.6% of tests.
Enterprise throughput claims
Cohere says the system is built for high-throughput serving in production, optimized around vLLM to handle high-volume speech workloads even as audio length varies, and that runtime and model-stack updates yielded up to 2x higher throughput. The company reports an RTFx (real-time factor multiple) score of 525, versus 146 for Whisper Large V3 and 66 for OmniASR 7B-LLM.
Key facts
- Cohere Transcribe Arabic is open source under the Apache 2.0 license, with weights on Hugging Face and hosted access via the Cohere API or Model Vault.
- Cohere reports a word error rate of 25.87 on the Hugging Face Arabic ASR Leaderboard, a 2.45-point improvement over Meta's OmniASR-LLM-7B and an 11-point improvement over Whisper Large V3.
- Cohere says the model ranks first on four of six composite task sets on the Open Universal Arabic ASR leaderboard, covering MSA, Egyptian, Gulf, Levantine, and Maghrebi dialects.
- In human evaluations, Cohere reports native Arabic speakers preferred the model over Whisper in 95.8% of head-to-head tests.
- Cohere reports an RTFx score of 525 for Transcribe Arabic, versus 146 for Whisper Large V3 and 66 for OmniASR 7B-LLM.
Check your understanding
What word error rate does Cohere report for Transcribe Arabic on the Hugging Face Arabic ASR Leaderboard?
Which two models does Cohere compare Transcribe Arabic against on the RTFx throughput measure?