Cohere has introduced Cohere Transcribe, an open-source automatic speech recognition model designed for enterprise transcription and speech-based AI workloads. Announced on March 26, 2026, the model is available for download and can also run through Cohere’s managed Model Vault platform. Cohere said the system was built from scratch with a focus on lowering word error rates while remaining practical for production use.
The company positioned the release for applications such as meeting transcription, speech analytics, and real-time customer support. Cohere Transcribe uses a conformer-based encoder-decoder architecture, has a 2-billion-parameter model size, and supports 14 languages, including English, French, German, Spanish, Japanese, Korean, Arabic, and Mandarin Chinese. The model is released under the Apache 2.0 license and is intended for local, GPU, and enterprise deployments.
Cohere said the model currently ranks first on Hugging Face’s Open ASR Leaderboard for accuracy, posting an average word error rate of 5.42%, ahead of several dedicated open- and closed-source alternatives in its benchmark comparison. The company also said human evaluations showed strong real-world transcript quality and that the model delivered high throughput suitable for production environments.In the article, the Cohere Team stated that speech is quickly emerging as a core modality for AI-driven workloads and automations, and described the launch as its initial step in bringing high-performance speech recognition into enterprise AI workflows.