Models

Reson8 Open-Sources Multilingual Speech Model to Challenge English-Centric AI

The Amsterdam startup has released Rumble-2-Flash, a 500 million-parameter transcription model supporting over 40 languages, as a free open-weights download to address the gap in non-English speech recognition.

·3 min read
Reson8 releases Rumble-2-Flash speech model as a free open-weights download
Reson8 releases Rumble-2-Flash speech model as a free open-weights download

Resonate Labs B.V., a speech recognition firm based in Amsterdam, unveiled Rumble-2-Flash today as a freely available model with open weights. The transcription system, built around 500 million parameters, handles more than 40 languages and targets developers and researchers seeking speech recognition capabilities beyond English.

The company contends that dominant speech models concentrate their training on English despite the fact that roughly one in 20 people use it as their native language. Accuracy deteriorates significantly when processing non-English speech, particularly with specialized vocabulary like pharmaceutical names. Reson8 indicates its technology powers meeting transcription tools, AI-driven receptionists, voice-based customer service systems, medical documentation, and accessibility features.

Generic speech models are built for English and make noticeably more mistakes in other languages. Languages like Dutch, Swedish and German are spoken by millions, yet they're an afterthought.

Raoul Ritter, co-founder and Chief Executive

Ritter founded the company alongside Thomas Kluiters and Jarno Verhagen. Reson8 positions Rumble-2-Flash among the most compact offerings listed on Hugging Face Inc.'s Open ASR Leaderboard. Testing across seven languages—including German, French and Dutch—yielded a mean word error rate of 7.09%, while English-only performance reached approximately 5%.

When evaluated on 10 lower-resource languages such as Maltese and Basque, the model achieved a mean word error rate of 17.6%. By comparison, OpenAI Group PBC's Whisper Large V3 registered 39.82% on the same assessment, per Reson8's figures.

Beyond standardized benchmarks, Reson8's technology incorporates training for challenging acoustic environments, including degraded phone-line recordings. According to Kluiters, who leads the company's AI efforts, "The real world doesn't happen in a recording studio."

The open-weights launch arrives following the model's November 3 introduction through Reson8's application programming interface, which preceded the company's rollout of its primary Rumble-2 offering by a week. The paid tier includes custom vocabulary functionality permitting users to incorporate up to 50,000 terms without retraining—a feature limited to Rumble-2.

Reson8 operates its infrastructure on graphics processing units housed in the Netherlands and maintains a no-retention policy for audio files and transcriptions. The company holds ISO 27001 certification and adheres to both the European Union's General Data Protection Regulation and its AI Act.

Rumble-2-Flash is distributed at no cost under an Apache 2.0 license. The model is also obtainable via Reson8's API beginning at €0.01 ($0.001) per hour of processed audio.

The startup secured €5 million ($5.5 million) in pre-seed capital during March, with Balderton Capital anchoring the investment and NP-Hard Ventures contributing. James Wise, a Balderton partner, remarked that speech is "quickly becoming a default interface across software."