Reson8 releases Rumble-2-Flash speech model as a free open-weights download

SiliconANGLE
Read full story· 4 min
Resonate Labs B.V. released Rumble-2-Flash, a 500M-parameter multilingual speech recognition model supporting 40+ languages, as a free open-weights download. It offers lower error rates than OpenAI's Whisper on low-resource languages and targets non-English transcription needs.

What happened

Resonate Labs B.V., based in Amsterdam, released its Rumble-2-Flash speech recognition model as a free download with open weights. The model, which has 500 million parameters, supports transcription in over 40 languages and is aimed at developers and researchers working with non-English audio. The release is available under an Apache 2.0 license and can also be accessed through the company's API. The model was previously available through Reson8’s API since November 3.

Key details

  • Rumble-2-Flash supports more than 40 languages and has 500 million parameters.
  • The model averaged a 7.09% word error rate across seven languages in multilingual tests.
  • On a test of 10 low-resource languages, it had a 17.6% mean word error rate, compared to 39.82% for OpenAI’s Whisper Large V3.
  • Custom vocabulary support is only available in Reson8’s Rumble-2 model, not Rumble-2-Flash.
  • The model is free under Apache 2.0 and API access starts at €0.01 ($0.001) per hour of audio.

Why it matters

Developers and organizations needing accurate speech recognition for languages other than English now have free access to a model that performs better on low-resource languages than some leading alternatives. This reduces transcription costs and increases access for users in multilingual and non-English environments. Companies and researchers working with sensitive data also benefit from the model’s compliance with European data protection and AI regulations.


More in Machine Learning