On March 26, 2026, Cohere released and open-sourced Transcribe (model identifier cohere-transcribe-03-2026): a 2-billion-parameter speech recognition model built on a Conformer encoder and Transformer decoder, published on Hugging Face under Apache 2.0. On release day it took first place on the Hugging Face Open ASR Leaderboard with a 5.42% average word error rate (WER).
The positioning is unambiguous: this is not a general-purpose voice assistant model. Transcribe targets business audio — recordings, contact-center calls, meetings — and converts it into text that can be searched, analyzed, and automated, with an explicit emphasis on running entirely on your own hardware.
Model Design: The 2B Trade-off
Transcribe covers 14 languages: nine European (English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish), four Asia-Pacific (Mandarin, Japanese, Korean, Vietnamese), and Arabic. Cohere’s blog post documents training data, model cards, and evaluation methodology, and stresses that the compact 2B footprint lets the model run on standard consumer hardware — a second signal, alongside Apache 2.0, that local deployment is the intended path.
The size is a deliberate choice: staying within single-card territory trades some ceiling for deployment flexibility, the opposite bet from closed models that only exist behind cloud APIs.
5.42% WER: Open ASR Enters a Photo Finish
On the Hugging Face Open ASR Leaderboard (the version Cohere cites, dated March 26, 2026), Transcribe’s 5.42% average WER leads a crowded field: Zoom Scribe v1 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%, and Whisper Large v3 at 7.44%.
The striking part is how tight the front pack is: first place and sixth place are separated by well under half a percentage point. Open-source ASR accuracy has entered a photo finish, and WER alone no longer separates products — differentiation is shifting to throughput, latency, and deployment sovereignty.
Human Evals and Throughput
Beyond WER, Cohere ran human preference evaluations: on average, evaluators preferred Transcribe’s output 61% of the time. Broken down by rival, the widest margin is over IBM Granite 4.0 (78%), then Whisper Large v3 (64%), Qwen3-ASR (55%), and just 51% against ElevenLabs Scribe v2 — effectively a coin flip, a reminder that at the top of the field the correlation between benchmark deltas and human judgment remains loose.
On throughput, Cohere claims best-in-class RTFx (the ratio of processing speed to audio duration) among open models above 1B parameters. For enterprises batch-digesting thousands of hours of recordings, that metric often determines total cost more than WER does.
Three Deployment Paths
The release deliberately ships three ways to run the model: download the weights from Hugging Face and run locally or on edge hardware; use Cohere’s free API (rate-limited); or go through Model Vault for production deployment and fine-tuning. Cohere’s recommended use cases include recording apps, contact centers, meeting systems, and regulated industries like finance and healthcare where data compliance constrains what can leave the premises.
TechCrunch’s same-day coverage centered on the “open-source voice model built specifically for transcription” framing; MarkTechPost read it through an enterprise-speech-intelligence lens, arguing transcription is the entry point for turning accumulated corporate audio into usable data. Cohere’s roadmap adds more languages, streaming and batch processing APIs, and domain-tuning techniques.
What It Means for Speech Builders
Three practical effects. First, the open ASR crown has changed hands again, and Apache 2.0 carries no extra usage restrictions — commercial products can adopt it without a licensing negotiation. Second, the local path is a hard requirement for data-residency-constrained workloads (call recordings, medical conversations, internal meetings), and Transcribe’s size targets exactly that segment. Third, the free-API-plus-Model-Vault combination shows open-model companies treating the model as the top of a product funnel while monetization happens at the deployment and platform layer — for developers, the weights are easy to adopt, but the surrounding workflow is where lock-in actually lives. For teams already standardized on Whisper pipelines, the migration cost is mostly validation time — rerunning accuracy checks on your own audio — since the weights, model card, and evaluation methodology are all published.
Sources
- Cohere Transcribe: Open-Source Speech Recognition — Cohere
- Cohere launches an open-source voice model specifically for transcription — TechCrunch
- Cohere AI Releases Cohere Transcribe: A SOTA ASR Model — MarkTechPost
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.