Cohere launches open-source 2B ASR model Transcribe

On March 31, 2026, Hacker News users pushed Cohere’s Transcribe launch to 154 points and 49 comments. The post stood out because it was not another general-purpose multimodal release. Instead, Cohere shipped a dedicated automatic speech recognition model with open weights and a clear production pitch.

In the official launch note, Cohere describes Transcribe as a 2B Conformer-based encoder-decoder trained from scratch. The model supports 14 languages including English, Japanese, Korean, Mandarin, Arabic, and several European languages, and it is released under Apache 2.0. Cohere also says the model currently ranks first on the Hugging Face Open ASR Leaderboard with an average word error rate of 5.42.

Why the release matters

• It is a purpose-built ASR model rather than a speech feature bolted onto a general assistant.
• Open weights and Apache 2.0 licensing lower friction for self-hosted enterprise deployments.
• Fourteen-language coverage makes the model relevant for meeting transcription, speech analytics, and support workflows.
• Cohere is offering the model through Hugging Face, its API, and Model Vault, which gives teams multiple deployment paths.

Cohere is positioning Transcribe as infrastructure for enterprise speech workflows rather than a demo model. The company highlights both local or private deployment and managed access through Model Vault and its API. It also pairs benchmark tables with human evaluation results, arguing that the accuracy gains survive beyond standardized datasets and into messy real-world audio.

The main qualification is that the leaderboard framing, throughput plots, and human preference data all come from Cohere’s own announcement. Buyers will still need independent latency testing and domain-specific evaluation before swapping out existing ASR pipelines. Even so, open weights, multilingual coverage, and a practical license make Transcribe one of the more concrete speech releases of late March 2026.

Community source: Hacker News discussion. Primary source: Cohere blog.