Cohere Releases Open-Weight ASR Model 'Transcribe' with 5.4% Word Error Rate, Apache-2.0 Licensed

2026-03-31T12:51:40.752Z·2 min read
Cohere has released Transcribe, an open-weight automatic speech recognition (ASR) model that achieves a 5.42% word error rate (WER), outperforming OpenAI's Whisper Large v3 and ElevenLabs Scribe v2.

Cohere has released Transcribe, an open-weight automatic speech recognition (ASR) model that achieves a 5.42% word error rate (WER), outperforming OpenAI's Whisper Large v3 and ElevenLabs Scribe v2.

Key Specifications

SpecificationDetail
Model namecohere-transcribe-03-2026
Parameters2 billion
LicenseApache-2.0 (commercial use OK)
WER (avg)5.42%
Languages14 (English, French, German, Italian, Spanish, Chinese, Japanese, Korean, Arabic, etc.)
DeploymentSelf-hosted on local GPU infrastructure

Performance Comparison

ModelWER
Cohere Transcribe5.42%
ElevenLabs Scribe v25.83%
Qwen3-ASR-1.7B5.76%
OpenAI Whisper Large v37.44%

Why This Matters

The ASR market has been dominated by closed APIs (OpenAI, ElevenLabs, Google) that offer accuracy but create data residency risks and lock users into external services. Cohere's Transcribe:

Use Cases

Analysis

Transcribe is significant because it closes the gap between open-source and proprietary ASR. At 2B parameters with Apache-2.0 licensing, it's practical for production deployment on commodity hardware. For organizations building AI agent pipelines that process audio, Transcribe offers a path to high-accuracy transcription without the latency, cost, and data privacy concerns of cloud APIs.

The 14-language support including Chinese, Japanese, and Korean makes it relevant for the Asian enterprise market, where data residency requirements are often stricter.

← Previous: Nvidia-Backed ThinkLabs AI Raises $28M to Tackle Power Grid Crunch with Physics-Informed AINext: Softr Launches AI-Native Platform: Moving Beyond 'Vibe Coding' to Production Business Apps →
Comments0