News · 2026-07-11

Cohere open-sources an Arabic speech model that beats Whisper and handles dialects and code-switching

Cohere open-sourced an Arabic speech-recognition model that it says is the most accurate open Arabic transcriber built to date -- and unlike most, it is designed for how Arabic is actually spoken, across regional dialects and mid-sentence switches into English. Released under the permissive Apache 2.0 license, it cuts word error rate well below OpenAI's widely used Whisper Large V3, and in Cohere's own human evaluation it was preferred over Whisper in about 96% of tests.

Key facts

• Accuracy: average word error rate about 25.9 on the Hugging Face Arabic ASR leaderboard -- lowest of any open model; Whisper Large V3 is about 36.9 (lower is better).
• Coverage: major dialect families (Egyptian, Gulf, Levantine, Maghrebi) plus Arabic-English code-switching.
• License: Apache 2.0, based on Cohere's 2-billion-parameter ASR model.
• Primary source: Cohere's Transcribe Arabic announcement.

To see why this is hard, you have to understand what makes Arabic tough for machines to transcribe. 'Arabic' is not one spoken language -- it is a formal written standard (Modern Standard Arabic) plus a family of everyday dialects, from Egyptian to Gulf to Levantine to Maghrebi, that differ enough that speakers from opposite ends of the region can struggle to understand each other. Most speech-recognition systems are trained heavily on the formal standard, so they stumble on the way people actually talk. On top of that, real Arabic conversation constantly code-switches -- dropping English words and phrases into an Arabic sentence -- which trips up models that expect one language at a time. Word error rate, the standard measure, counts what fraction of words a transcriber gets wrong; lower is better, and on messy real-world Arabic the numbers for general models get ugly fast.

Cohere's model attacks exactly that messiness. It posts an average word error rate around 25.9 on the Hugging Face Arabic leaderboard -- the best of any open model -- against roughly 28.3 for Meta's OmniASR and about 36.9 for Whisper Large V3. The gap widens on the hardest dialect test sets: on one spontaneous-speech benchmark it more than halves Whisper's error rate. And in a head-to-head human evaluation, listeners preferred Cohere's transcriptions over Whisper's in about 96% of tests, and even preferred it for English spoken with a heavy Arabic accent about 77% of the time. Cohere frames the release as 'a proud advance in the region's sovereign AI capabilities, bringing frontier performance to millions of Arabic-speakers' -- part of a broader 'sovereign AI' pitch about regions owning their own language technology.

Why it matters is that speech recognition is infrastructure, and for a language spoken by hundreds of millions of people, open and accurate infrastructure has been missing. A closed API that half-understands your dialect is a poor foundation for building voice assistants, captioning, call-center tools, or accessibility software; open weights you can self-host and fine-tune are a real one. This release bundles naturally with the week's other open-ecosystem news -- Alibaba's Qwen3Guard safety model -- as evidence that the open world is now filling in the unglamorous pieces (non-English speech, safety filtering) that proprietary systems have quietly monopolized.

The honest caveat is that the standout numbers -- especially the 96% human-preference figure -- come from Cohere's own evaluation, so they should be read as the vendor's benchmark rather than an independent one. The word-error-rate results sit on a public leaderboard others can reproduce, which is stronger evidence, but 'best open model' is a claim that only holds until the next release. What is not in dispute is the direction: open Arabic speech recognition just took a real step forward, under a license that lets anyone build on it.

Primary source, verified: read the paper →

Ground Truth. (2026, July 11). Cohere open-sources an Arabic speech model that beats Whisper and handles dialects and code-switching. Ground Truth. https://groundtruth.day/news/cohere-transcribe-arabic-open-source-speech.html

@misc{groundtruth:cohere-transcribe-arabic-open-source-speech,
  title  = {Cohere open-sources an Arabic speech model that beats Whisper and handles dialects and code-switching},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/cohere-transcribe-arabic-open-source-speech.html}
}

Topics: speech-recognition · open-weights · cohere · arabic · multilingual