Cohere has released its new open-source Automatic Speech Recognition (ASR) model, Transcribe Arabic, which is designed to capture the nuances and dialectical richness of Arabic speech. The model is built for enterprise applications and developer use, supporting dialect variation, bilingual Arabic-English speech, and domain-specific vocabulary. It is the most accurate open-source Arabic speech-to-text model to date, with a word error rate (WER) of 25.87 on the Hugging Face Arabic ASR Leaderboard. According to Cohere, the model delivers substantial gains over its predecessor, Cohere Transcribe, across both Arabic and bilingual Arabic-English audio. Developers can access the model under the Apache 2.0 license via Hugging Face or the Cohere API. Source: cohere

The release of Transcribe Arabic marks a significant step in advancing Arabic language technology, which has historically been under-served by state-of-the-art AI systems. Arabic is spoken by over 300 million people across more than 30 dialects, each shaped by distinct cultural and historical contexts. The diversity of Arabic speech presents challenges for natural language processing, particularly in maintaining dialectal nuance while ensuring broad applicability. Cohere addressed this by training its model extensively on data spanning Arabic dialects, professional language, and varied acoustic conditions. The result is a new state-of-the-art solution for capturing Arabic speech, ready for production use and openly available to all. Source: cohere

Cohere Transcribe Arabic achieved the lowest average 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 Meta’s OmniASR-LLM-7B and an 11-point improvement over OpenAI’s Whisper Large V3. Human evaluators preferred Cohere Transcribe Arabic over Whisper in 95.8% of tests, and found it broadly comparable to Whisper on English spoken with an Arabic accent, with Cohere Transcribe Arabic preferred in 52.6% of tests. The model also scored highest on dialect faithfulness and robustness to code-switching compared to other models. Source: cohere