Reson8 releases 0.5B-parameter speech model supporting 40 languages

What happened
Reson8, a speech recognition startup based in Amsterdam, released its Rumble-2-Flash model as a free open-source download. The model contains 500 million parameters and supports automatic transcription in over 40 languages. It is designed to run quickly on standard frameworks and is available under an Apache 2.0 license. The release follows the model's earlier debut in Reson8's API and is aimed at developers and researchers working with non-English audio.
Key details
- Rumble-2-Flash averages a 7.09% word error rate across seven major languages and 17.6% on ten low-resource languages.
- The model outperformed OpenAI's Whisper Large V3 on low-resource language benchmarks, with Whisper scoring 39.82% error rate on the same test.
- Custom vocabulary support is only available in Reson8's flagship Rumble-2 model, not in Rumble-2-Flash.
- The model is available for free download and through Reson8's API, with API pricing starting at €0.01 ($0.001) per hour of audio.
Why it matters
Developers and organizations working with speech data in languages other than English now have free access to a high-accuracy, multilingual speech recognition model. This reduces costs for those needing transcription in less commonly supported languages and offers better performance on low-resource languages than some larger commercial models. Companies and researchers can integrate the model into their applications without licensing fees, and can process audio securely under European data protection standards.
