Cohere launches high accuracy Arabic transcription model
Open-source model tops Arabic ASR leaderboard, beats Whisper for Arabic
#Global #LLMs - Toronto-headquartered AI company Cohere has released Cohere Transcribe Arabic, an open-source speech-to-text model it says is the most accurate available for Arabic. The 2-billion-parameter model scores a word error rate of 25.87 on the Hugging Face Arabic ASR Leaderboard, and native Arabic-speaking reviewers preferred its transcripts over Whisper’s in 95.8% of tests. According to the developer, eating Meta’s OmniASR-LLM-7B and OpenAI’s Whisper Large V3 across dialects. Transcribe Arabic is released under the Apache 2.0 licence.
SO WHAT? - Arabic language AI has lagged behind English, in particular when it comes to speech recognition, text-to-speech and transcription. Spoken natively by more than 300 million people across the Middle East and North Africa, the language has roughly 30 dialects, but only one written form (Modern Standard Arabic or MSA). The prevalence of dialects makes AI voice all the more important commercially, as large service providers seek to engage their audiences in their native dialects. Existing tools tend to flatten dialects into Modern Standard Arabic and stumble on Arabic-English code-switching. Cohere built Transcribe Arabic to handle both well, and its open source availability allows developers to build sovereign AI systems.
KEY POINTS:
Cohere released Cohere Transcribe Arabic as an open-source, 2-billion-parameter automatic speech recognition model, built on the architecture of its Cohere Transcribe model launched in March.
The model achieves a word error rate of 25.87 on the Open Universal Arabic ASR Leaderboard, beating Meta’s OmniASR-LLM-7B (28.32) and OpenAI’s Whisper Large V3 (36.86).
Cohere Transcribe Arabic ranks first on four of six composite test sets spanning Modern Standard Arabic, Egyptian, Gulf, Levantine and Maghrebi dialects. The biggest gains were found on conversational, multi-dialect datasets like Casablanca.
According to Cohere, native Arabic-speaking evaluators preferred its transcripts over Whisper’s in 95.8% of head-to-head tests, rating it highest on accuracy, dialect faithfulness and handling of Arabic-English code-switching.
The model preserves regional phrasing and workplace vocabulary, such as keeping terms like “annual leave” and “HRIS” intact, rather than mistranslating or garbling them, as rival models did in test transcripts.
For English spoken with an Arabic accent, human evaluators preferred it over Cohere’s original Transcribe model in 77.2% of tests, and rated it roughly on par with Whisper in 52.6% of tests.
The model runs on consumer hardware without requiring cloud services, and is optimised around vLLM for high-throughput production use, hitting a real-time processing speed multiple of 525, against 146 for Whisper and 66 for OmniASR.
Cohere Transcribe Arabic is available now on Hugging Face, through the Cohere API with free rate-limited access, or via a paid, dedicated Model Vault deployment for production use without limits.
Cohere has raised more than $1.6 billion from investors including Nvidia, AMD Ventures, Salesforce Ventures, Oracle and Cisco, and counts AI researchers Geoffrey Hinton and Fei-Fei Li among its backers.
ZOOM OUT - The Transcribe Arabic model has been announced during the same week as a deal between Cohere and Saudi Arabia’s national AI company HUMAIN. The company has, until now, built its Arabic AI capability around the Kingdom’s first sovereign AI model ALLaM, first launched by SDAIA in 2023. The new deal will see Cohere and HUMAIN develop Arabic AI models jointly, alongside special-purpose models built for enterprise and sector-specific use cases. The deal is the first significant move by HUMAIN to partner with a large language model developer to build sovereign AI models for Saudi Arabia.
LINKS
Cohere Transcribe Arabic code (Hugging Face)
Transcribe Arabic technical blog (Cohere)
Transcribe Arabic technical blog (Hugging Face)
Arabic ASR Leaderboard (Hugging Face)
Read more about Arabic-first AI models:
TokenAI’s Horus Hiero multimodal AI reads hieroglyphics (Middle East AI News)
Inception42 advances Arabic AI for enterprise with Seraj (Middle East AI News)
Stanford, Arabic.AI launch Arabic Enterprise AI benchmark (Middle East AI News)
CNTXT AI acquires Actualize to expand Arabic voice agents (Middle East AI News)
Native Arabic AI application builder launched in UAE (Middle East AI News)