Moonshine: Open STT Models Aim to Beat Whisper

Resemble AI has released Moonshine, a family of open-weight speech-to-text models that the company claims achieve higher accuracy than OpenAI's Whisper Large v3. The models are published under the permissive MIT license, making them straightforward to adopt in commercial as well as research projects, and are available on GitHub.

Moonshine arrives in the under-1B parameter range, positioning it as a comparatively lightweight option in a space where Whisper has become the default reference point for open transcription. The compact size matters: smaller automatic speech recognition models are easier to run on edge devices and consumer hardware, and they reduce the cost and latency of real-time transcription.

Why it matters

Whisper has dominated open speech recognition since its release, so any credible challenger draws attention. A permissively licensed alternative that claims to edge out Whisper Large v3 on accuracy could give developers a meaningful choice, particularly for applications where model footprint and licensing terms are deciding factors.

• MIT license, suitable for commercial use
• Open weights in the sub-1B parameter class
• Positioned as a direct accuracy competitor to Whisper Large v3

As always, vendor accuracy claims warrant independent verification across languages, accents, and noisy conditions before teams commit to a swap. But the combination of open weights, a permissive license, and a small footprint makes Moonshine worth a look for anyone building transcription into their stack.