Summary: Cohere open-sourced Transcribe on March 26, 2026 as a 2-billion-parameter ASR model under Apache 2.0, topping the Hugging Face Open ASR Leaderboard with 5.42% average word error rate. The model supports 14 languages, runs on consumer hardware, and targets recordings, contact centers, meetings, and regulated industries. The article notes that deployment, throughput, and data control now differentiate top models more than small WER gaps.
Cohere opened up its speech recognition model Transcribe (identifier cohere-transcribe-03-2026) on March 26, 2026. Inside is a Conformer encoder paired with a Transformer decoder and a total of 2 billion parameters, with the weights sitting on Hugging Face under the Apache 2.0 license. That same day, the model took over the number-one position on the Hugging Face Open ASR Leaderboard with an average word error rate (WER) of 5.42%.
There is no ambiguity about what this model is for. It is not a general-purpose voice assistant; Transcribe is aimed at business audio from the start — recordings, contact-center calls, and meetings turned into text that can be searched, analyzed, and automated — and it is explicitly built to run entirely on your own hardware.
Model Design: What 2B Buys, and What It Gives Up
Fourteen languages are covered: nine European ones (English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish), four from Asia-Pacific (Mandarin, Japanese, Korean, Vietnamese), and Arabic on top of that. The blog post accompanying the release spells out the training data, model cards, and evaluation methodology, and notes that the compact 2B footprint runs on standard consumer hardware — after Apache 2.0, that is the second clear signal that local deployment is the intended path.
That size was a conscious decision. Holding the model within single-card reach gives up a little headroom in exchange for deployment freedom — the exact opposite bet from closed models that live only behind cloud APIs.
5.42% WER: A Photo-Finish Race in Open ASR
On the Hugging Face Open ASR Leaderboard (using the version Cohere cites, current through March 26, 2026), the 5.42% average WER puts Transcribe at the front of a deep field: Zoom Scribe v1 follows at 5.47%, IBM Granite 4.0 1B Speech at 5.52%, NVIDIA Canary Qwen 2.5B at 5.63%, Qwen3-ASR-1.7B at 5.76%, ElevenLabs Scribe v2 at 5.83%, with Whisper Large v3 further back at 7.44%.
How tightly packed the leaders are is the real story: going by the numbers Cohere cites, the first six places sit within less than 0.45 percentage points of one another. Accuracy among open-source ASR models has turned into a photo finish where WER alone can no longer set products apart, and the contest is moving to throughput, latency, and deployment sovereignty.
Human Preference Scores and Throughput
WER is not the only yardstick: Cohere also ran human preference evaluations, with evaluators favoring Transcribe’s output 61% of the time on average. Split by opponent, the biggest gap comes against IBM Granite 4.0 at 78%, next is Whisper Large v3 at 64%, then Qwen3-ASR at 55%, and against ElevenLabs Scribe v2 the edge shrinks to 51% — as close to a coin flip as it gets. The reminder here is that among top-tier models, benchmark deltas and what human listeners actually perceive still track each other only loosely.
As for throughput, Cohere asserts that RTFx — processing speed measured against audio duration — is the best among open models above 1B parameters. When an enterprise has to chew through thousands of hours of recordings in batches, that figure tends to shape total cost more than WER does.
Three Ways to Run It
Three routes ship with the release on purpose: take the weights from Hugging Face and run them locally or on edge devices; connect to Cohere’s free API, which carries rate limits; or go through Model Vault for production-grade deployment and fine-tuning. The use cases Cohere recommends cover recording apps, contact centers, meeting systems, and regulated industries such as finance and healthcare, where data compliance constrains what can leave the premises.
Day-of coverage split along those lines: TechCrunch framed its story around the “open-source voice model built specifically for transcription” positioning, while MarkTechPost looked at the release through an enterprise-speech-intelligence lens, contending that transcription is the entry point where accumulated corporate audio starts becoming usable data. Further out, Cohere’s roadmap adds more languages, streaming and batch processing APIs, and domain-tuning techniques.
What Speech Builders Should Take Away
Three practical effects follow. For one, the open ASR crown has switched heads yet again, and because Apache 2.0 adds no usage restrictions of its own, commercial products can adopt it straight away. Next, for workloads where data cannot freely leave the premises (call recordings, medical conversations, internal meetings), the local path is a hard requirement, and Transcribe’s sizing is aimed right at that need. Last, the free-API-plus-Model-Vault pairing shows open-model companies treating the model as the top of a product funnel while monetization happens at the deployment and platform layer — picking up the weights is easy, but the surrounding workflow is where lock-in actually lives. Teams already standardized on Whisper pipelines will find the migration cost is mostly validation time — rerunning accuracy checks on their own audio — given that the weights, model card, and evaluation methodology are all published.