Cohere releases Transcribe Arabic, an open-source Arabic speech recognition model

Cohere has open-sourced Cohere Transcribe Arabic, a 2B-parameter automatic speech recognition model purpose-built for Arabic dialects, code-switching, and business/developer speech, releasing the weights under an Apache 2.0 license.

What's new

Cohere describes the release as "the world's most accurate open-source model for Arabic speech recognition." The model is a specialized version of Cohere's earlier Transcribe ASR system, adapted specifically for the realities of spoken Arabic: heavy dialect variation across regions, bilingual Arabic-English speech, mid-sentence code-switching, and domain-specific vocabulary common in business and developer settings.

On accuracy, Cohere states the model "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," and that in head-to-head evaluations "it was preferred over Whisper in 95.8% of tests" — a direct comparison against OpenAI's widely used open Whisper models, which have long been a default choice for open Arabic transcription despite known weaknesses on dialect and code-switching.

On availability, Cohere says: "Cohere Transcribe Arabic is available under the Apache 2.0 license. Developers can download the weights and read our quickstart implementations on Hugging Face, or access the hosted model through the Cohere API or Model Vault." That gives developers three paths in: self-hosted open weights, a Hugging Face-hosted space for quick testing, or Cohere's own managed API.

Context

Arabic speech recognition has lagged well behind English in open-source tooling, largely because dialectal variation across the Arabic-speaking world (Gulf, Levantine, Egyptian, Maghrebi, and Modern Standard Arabic) makes a single model's performance highly inconsistent across regions and speakers. Whisper, despite broad multilingual coverage, has been reported to struggle specifically with this variation and with code-switching between Arabic and English in professional contexts — the exact gap Cohere is targeting. The release also lands alongside a broader Cohere push into Arabic-language AI, including a reported infrastructure partnership in Saudi Arabia focused on sovereign Arabic models, positioning speech as one more layer of that regional strategy.

Why it matters

Open-weight, Apache-licensed releases lower the barrier for regional developers, government agencies, and enterprises across the Middle East and North Africa to build Arabic voice products without depending on a closed API from a US-based provider — relevant for call centers, media transcription, and voice interfaces serving the roughly 400 million Arabic speakers worldwide. A credible, independently-verifiable win over Whisper on a public leaderboard also signals that dialect-and-code-switching-aware fine-tuning, rather than raw model scale, is where meaningful ASR gains are now coming from for underserved languages — a pattern likely to repeat for other high-speaker-count, dialect-heavy languages that general-purpose models handle poorly.