Cohere has released Cohere Transcribe, an open-source automatic speech recognition (ASR) model aimed at the growing enterprise demand for embedding transcription directly into applications. The model has roughly 2 billion parameters and is small enough to run on consumer-grade GPUs and at the edge.
Announced on March 26, 2026, Cohere Transcribe is released under the permissive Apache 2.0 license and supports 14 languages. At launch it ranked first on Hugging Face’s Open ASR leaderboard, with a reported average word error rate of about 5.42%, placing it ahead of widely used systems such as OpenAI’s Whisper. The company has said it plans to integrate the model into North, its enterprise AI agent platform.
How speech recognition has evolved
Cohere Transcribe reflects a broader shift in ASR design. Earlier speech systems relied on deep-learning techniques such as long short-term memory recurrent neural networks and, later, transformer-based architectures. Larger models in particular often struggled to deliver low latency. Cohere Transcribe instead uses a conformer-based encoder-decoder architecture, which the company says allows high throughput while remaining compact enough for on-device deployment. A related overview of running such systems locally appears in this guide to real-time speech models.
A crowded and expanding market
As infrastructure and model quality have matured, ASR has spread into customer service, banking, sales, and marketing, drawing entrants ranging from established vendors such as IBM and Alibaba to video-conferencing providers. Zoom, for example, has added AI-powered real-time translation features that let participants follow a conversation in their own language. The common thread is a move toward smaller models that can run closer to the user rather than only in the cloud.
Speech remains central to how people interact with AI systems, a point underscored by industry analysts who note that voice interfaces helped popularize AI in the first place. Lian Jye Su, an analyst at Omdia, has highlighted Cohere Transcribe’s small size and open-source release as notable: making a model openly available invites developers to test it, build on it, and potentially adopt a commercial version later — a strategy that companies including Meta have used to broaden the reach of their models.
Why open source matters here
Cohere has historically focused on text generation and enterprise language models, so an ASR release marks a move into adjacent territory. The opportunity, analysts suggest, lies with enterprises looking to replace older transformer-based speech systems with a new generation of smaller models suited to edge devices. By releasing the weights openly, Cohere positions Transcribe to gain traction with developers while leaving room to monetize through its API and platform integrations.
Limitations and what to watch
Benchmark leadership is a snapshot rather than a permanent ranking; ASR leaderboards change quickly as new models appear, and word-error-rate figures depend heavily on the test sets used. Real-world accuracy also varies with accents, background noise, domain-specific vocabulary, and audio quality, so leaderboard results do not guarantee equivalent performance on a given workload. As with any transcription system, outputs intended for records or decisions should be reviewed rather than treated as exact.
