Cohere launches Transcribe, a 2B Apache 2.0 speech recognition model

What Cohere announced

On March 26, 2026, Cohere introduced Transcribe on X as a new state-of-the-art open-source speech recognition model. The official release gives the claim more substance: Cohere says Transcribe is a 2B parameter Conformer-based encoder-decoder model, trained from scratch for production-grade automatic speech recognition rather than as a research demo.

What the release page says

According to Cohere, Transcribe supports 14 languages and ships under the Apache 2.0 license. The company says the model currently ranks first on Hugging Face’s Open ASR Leaderboard with an average word error rate of 5.42, ahead of other dedicated open and closed speech systems listed there. Cohere also emphasizes that the model is designed to balance low word error rate with high throughput, which matters more in real deployments than benchmark wins alone.

The release page positions Transcribe as a practical building block for meeting transcription, speech analytics, search over audio, and real-time customer support agents. Cohere is making it available three ways: as open weights on Hugging Face, through its API for experimentation, and inside Model Vault for managed private deployment. That mix targets both developers who want local infrastructure control and enterprises that want hosted operation without self-managing the stack.

Why it matters

Speech has been a relatively fragmented part of the AI stack, with strong models often gated behind commercial APIs or narrower licenses. Cohere is betting that an Apache-licensed, leaderboard-leading ASR model with a manageable serving footprint can pull speech recognition closer to the mainstream enterprise toolchain. If the latency and accuracy claims hold up outside the launch benchmarks, Transcribe could become a serious default for organizations that want open speech infrastructure without settling for second-tier quality.

Source: Cohere X post · Cohere release page