Cohere has launched Cohere Transcribe, a state-of-the-art automatic speech recognition model that tops the HuggingFace Open ASR Leaderboard with a 5.42% word error rate. The open-source Conformer-based system supports 14 languages and delivers production-ready performance for enterprise AI workflows. Trained from scratch with a focus on minimizing word error rate while maintaining practical deployment characteristics, the model represents a significant advance in speech-to-text technology.

Unlike research artifacts, Cohere Transcribe was designed for everyday use with a manageable 2B parameter footprint suitable for GPU and local deployment. The model achieves best-in-class serving efficiency while maintaining accuracy across diverse real-world conditions including multiple speaker environments, boardroom acoustics, and various accents. Human evaluations confirm the benchmark results, with trained reviewers consistently preferring Transcribe's transcription quality for accuracy, coherence, and usability across supported languages.

Available today via Hugging Face for local deployment or through Cohere's Model Vault for managed inference, Transcribe extends the Pareto frontier by combining state-of-the-art accuracy with exceptional throughput. Early partners report impressive speed, converting minutes of audio to usable transcripts in seconds. Cohere plans deeper integration with its North AI agent orchestration platform, evolving Transcribe from high-accuracy transcription into a broader foundation for enterprise speech intelligence.