AI BriefingArabic Speech Recognition
Key point
Cohere has released an open-source Arabic speech recognition (ASR) model with accuracy that surpasses Whisper.
Details
Cohere unveiled Cohere Transcribe Arabic, an open-source model that reflects the diverse dialects of Arabic and the particularities of business environments. This model is based on a 2B-parameter ASR model, and it maximizes the ability to handle dialect differences, Arabic-English code-switching, and specialized terminology.
Cohere Transcribe Arabic recorded an average Word Error Rate (WER) of 25.87 on Hugging Face's Arabic ASR leaderboard, significantly outperforming Meta's OmniASR-LLM-7B (28.32) and OpenAI's Whisper Large V3 (36.86). Notably, the model showed such high accuracy that human reviewers preferred it over Whisper 96% of the time.
Key features are as follows:
- Dialect and mixed-language support: Effectively handles the diverse dialects used by over 300 million Arabic speakers and Arabic-English code-switching situations.
- Enterprise-grade performance: Provides high accuracy optimized for business and developer environments.
- Open source availability: Released under the Apache 2.0 license, allowing developers to download the weights or use it via the Cohere API.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.