Cohere Transcribe

Cohere stepped into open speech recognition in March 2026 with cohere-transcribe-03-2026, a 2B-parameter model released under Apache-2.0 that posted a 5.42 average word error rate to lead the Open ASR Leaderboard at launch, with LibriSpeech clean down at 1.25 and solid numbers on harder sets like earnings calls and AMI meetings. The design is classic and efficient: a large Conformer encoder distills the acoustics, then a lightweight Transformer decoder emits tokens, keeping inference cheap relative to LLM-based transcribers. Coverage spans 14 languages across Europe (English, French, German, Spanish, Italian, Portuguese, Dutch, Polish, Greek), APAC (Mandarin, Japanese, Korean, Vietnamese), and Arabic. Weights ship as safetensors on Hugging Face with reference code for the Transformers library and a vLLM path for production serving, so it drops into existing pipelines with a few lines of Python. There is no official desktop app or CLI, making this a building block for developers rather than an end-user tool, and a single mid-range GPU handles it comfortably.