Cohere Transcribe Arabic tops Whisper ASR leaderboard

// 2h agoMODEL RELEASE

Cohere Transcribe Arabic tops Whisper ASR leaderboard

Cohere has released Cohere Transcribe Arabic, an open-source ASR model under the Apache 2.0 license that leads the Open Universal Arabic ASR Leaderboard. Reviewers preferred the new model over OpenAI's Whisper in 96% of human evaluation tests, citing its strong performance on code-switching and multiple dialects.

// ANALYSIS

Cohere's release of a highly specialized regional model is a strategic play to win over developer and enterprise markets in the Middle East where localized nuances are critical.

–**Outperforming Whisper:** Leading the ASR leaderboard and achieving a 96% human preference rate against Whisper establishes it as a formidable open-source alternative.
–**Dialect and Accent Mastery:** Explicit design for code-switching and multiple dialects solves a major friction point for real-life transcription.
–**Permissive Open-Source Licensing:** Using the Apache 2.0 license makes it highly appealing for enterprises concerned with sovereign AI and data privacy.

// TAGS

coheretranscribe-arabicsttarabicopen-sourceapache-2.0hugging-face

DISCOVERED

2h ago

2026-07-07

PUBLISHED

2h ago

2026-07-07

RELEVANCE

8/ 10

AUTHOR

cohere

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

LAUNCH51m ago

Runway launches Slide Maker presentation tool

Runway has launched Slide Maker, a generative tool that allows users to create custom presentation slides from text descriptions. Instead of using templates, the AI dynamically designs slide layouts and graphics to automate presentation design.

LAUNCH53m ago

Archil launches dedicated agent tools

Archil has launched pre-built agent tools to streamline connecting AI agents to its persistent cloud storage layer. This allows developers to integrate Archil's high-performance storage directly into agent sandboxes, enabling agents to read, write, and execute code on shared datasets using standard POSIX operations.

RESEARCH55m ago

Diffusion LLMs claim half of ICML 2026 awards

An analysis of ICML 2026 accepted papers shows that diffusion-based large language models (diffusion LLMs) are establishing a strong presence, accounting for 1.5% of accepted submissions (98 of 6,352 papers). While representing a minority architecture, they punch above their weight, making up 2.4% of oral presentations and securing 50% of the Outstanding Paper Awards.