AccessibilityThursday, March 26, 2026· 2 min read

Cohere open-sources a compact, self-hostable voice model for fast, private transcription

TL;DR

Cohere released an open-source voice model optimized for transcription that runs on consumer-grade GPUs and is lightweight at 2 billion parameters. Supporting 14 languages and designed for self-hosting, the model lowers barriers to private, affordable, and accessible speech-to-text.

Key Takeaways

  • 1Compact 2 billion-parameter model designed to run on consumer-grade GPUs for easy self-hosting.
  • 2Open-source release encourages community improvements and wider adoption.
  • 3Supports transcription in 14 languages, expanding multilingual accessibility.
  • 4Enables private, low-cost on-premise transcription for creators, organizations, and accessibility tools.

Cohere launches a practical, open voice model for transcription

Cohere has released a new open-source voice model specifically aimed at transcription use cases. At a relatively light 2 billion parameters, the model is designed to run on consumer-grade GPUs, making it practical for individuals and organizations that want to self-host speech-to-text capabilities without relying on cloud APIs.

The model currently supports 14 languages, bringing accurate and private transcription within reach for a broad set of users. By targeting a compact footprint, Cohere has prioritized accessibility and affordability—allowing independent developers, journalists, educators, and small businesses to deploy transcription locally with modest hardware.

Open-sourcing the model invites the community to contribute improvements, build custom fine-tuned variants, and integrate the technology into accessibility tools like captioning and assistive apps. The ability to self-host also addresses privacy and compliance concerns for sensitive audio data, while lowering operating costs compared with some cloud transcription services.

Overall, this release is a practical win for democratising speech-to-text: it combines multilingual support, manageable compute requirements, and an open development model to accelerate real-world transcription use-cases across many sectors.

Get AI Wins in Your Inbox

The best positive AI stories delivered to your inbox. No spam, unsubscribe anytime.