Enterprise AI company Cohere entered the voice market on March 26, releasing Transcribe — its first speech recognition model — under an Apache 2.0 license. It immediately topped the HuggingFace Open ASR Leaderboard with an average word error rate of 5.42%, outperforming Whisper Large v3, ElevenLabs Scribe v2, and every other open or closed-source dedicated ASR model on the benchmark.

For businesses thinking about voice AI, Cohere’s move matters for reasons beyond the benchmark score.

What Cohere Released

Transcribe (cohere-transcribe-03-2026) is a 2-billion-parameter conformer encoder-decoder model available for free on Hugging Face and via Cohere’s API at no cost. It supports 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese, Japanese, Korean, Vietnamese, and Arabic.

The standout technical specification is throughput: Cohere claims the model processes 525 minutes of audio per minute, a high figure for its parameter class. That matters for any business running transcription at volume — call centers, note-taking tools, document processing pipelines.

Because it runs on consumer-grade GPUs, businesses can self-host the model on their own infrastructure. That closes one of the most common objections to cloud-based transcription services: data leaving the organization.

Why Cohere Is Doing This

Cohere has been a text-first enterprise AI company since its founding. It built its name on large language models for business applications, not voice. So entering the ASR market is a shift in direction — and the reasoning is fairly transparent.

Cohere’s North agent orchestration platform handles complex, multi-step AI workflows. Voice is a natural input channel for those workflows. A customer service agent, an internal operations agent, a meeting summarization tool — all of them benefit from high-quality speech recognition at the front end. Building that capability in-house rather than depending on a third-party ASR provider is a strategic move toward a more complete enterprise AI stack.

The open-source release under Apache 2.0 is consistent with Cohere’s approach to developer relationships. Making Transcribe freely available builds goodwill in the developer community and positions Cohere as a serious player across the full enterprise AI stack, not just text.

The Benchmark Numbers

Transcribe leads the HuggingFace Open ASR Leaderboard at 5.42% average word error rate. Human evaluators rated its transcriptions for accuracy, coherence, and usability — Cohere claims an average 61% win rate over competing models in those assessments.

Throughput at 525 minutes of audio per minute means real-time processing with headroom to spare.

The limitations are worth noting: Transcribe scores a 44% win rate in German and 48% in both Spanish and Portuguese, falling behind rivals in three of its fourteen supported languages. It also benefits from a noise gate or voice activity detection layer in noisy environments, as the model will transcribe low-level background noise if given the chance.

What This Means for Business

For businesses evaluating voice AI: Cohere Transcribe adds a new credible option to a market that was already moving fast. The fact that it runs on-premises and is free to use removes two of the most common friction points in enterprise voice AI procurement: cost and data residency. If your security or legal team has concerns about sending audio to third-party cloud services, a self-hosted open-source model addresses that directly.

For businesses already using voice AI: The market just got more competitive. When a well-funded enterprise AI company enters ASR with a state-of-the-art open-source model, it puts pressure on every incumbent. Pricing from commercial providers may soften. Quality will improve. That is good for enterprise buyers.

For contact centers and high-volume transcription users: The throughput figure — 525 minutes of audio per minute — means Transcribe can handle serious production workloads. If you are running a call center that logs thousands of calls daily, the model’s cost (free at API, or self-hosted) and throughput make it worth a serious evaluation.

For developers building voice-first applications: Apache 2.0 licensing means you can integrate Transcribe into commercial applications without licensing constraints. That is a meaningful advantage over models with more restrictive terms.

The Bigger Voice AI Picture

Voice AI crossed $22 billion in market size in 2026, and 67% of Fortune 500 companies now run production voice AI systems. The market has matured past the “interesting experiment” phase into mission-critical infrastructure.

What Cohere’s entry signals is that the competitive pressure in voice AI is intensifying at the infrastructure layer. When you have IBM partnering with ElevenLabs for premium enterprise voice, Mistral releasing open-source voice models, and now Cohere entering ASR — the story is clear: voice AI is becoming table-stakes capability for any company serious about enterprise AI.

For businesses still on the sidelines, the window to make a considered decision is shorter than it was a year ago. The tools are available, the benchmarks are strong, and the economics have shifted dramatically in favor of deployment.

The Practical Question

The relevant question for most businesses is not whether speech recognition has gotten good enough — it clearly has. The question is whether you have thought through the workflow around it. Transcription is a data collection layer. The value comes from what you do with that data: surfacing insights, triggering follow-up actions, feeding into agent workflows.

Businesses that think about voice AI as an input layer for broader automation will extract more value than those treating it as a standalone transcription service.

---

The practical next step is the free Working With Claude field guide. Thirty-two pages covering the ecosystem, Claude Code, and how to govern a rollout properly. Get your copy.