Reson8 releases open speech recognition model Rumble-2-Flash
Amsterdam start-up Reson8 has released Rumble-2-Flash as a free download with open model weights. The model turns recorded speech in more than forty languages into text, including Dutch.
Rumble-2-Flash turns recorded speech into text in more than forty languages, including Dutch. The model has 500 million parameters, the adjustable numbers in which it stores what it has learned. That is small compared with the large speech models of tech companies. Reson8 released it on 18 November 2026 through Hugging Face, a popular website where AI models are shared.
Anyone may download and use the model weights, including in commercial software, because the model falls under the free Apache 2.0 licence. It works with common open software for running AI models. It recognises by itself which language is being spoken and can write out figures as numbers or as words.
Anyone who does not want to run the model themselves can use it through Reson8's API, a connection that lets other software call the service. That costs from €0.01 per hour of audio. The company runs this service on its own computers in the Netherlands and says it does not keep any recordings or transcripts.
Rumble-2-Flash is meant for audio that has already been recorded. The larger model Rumble-2 can tell speakers apart; according to the company, transcribing live conversations is also a task for a larger model.
How the small model learned
Reson8 made Rumble-2-Flash by passing on knowledge from Rumble-2. The small model transcribed recordings itself, after which the large model assessed every piece of that text. In this way the small model learned from its own mistakes. The two models cut text into pieces in different ways; a new method, GOLD, still makes those pieces comparable. The model is built on the design of Qwen3-ASR, an open speech model from Alibaba's Qwen team.
Results on speech tests
Speech recognition is measured by the percentage of words that go wrong, compared with a transcript made by a person. Such a score also depends on spelling conventions and on mistakes in that human transcript. On the multilingual part of the Open ASR Leaderboard, a public ranking on Hugging Face, Rumble-2-Flash scored an average of 6.96 per cent errors across nine languages, with 6.78 per cent for Dutch. According to Reson8, that is better than OpenAI's Whisper large-v3 and at least as good as Meta's omniASR-LLM 7B, a model fifteen times its size.
On a test by the company AppTek with 302 hours of real calls to telephone helplines, in fourteen English accents, 7.86 per cent of the words went wrong. Nvidia's Nemotron 3.5 Transcribe, a model sixty times its size, scored 7.78 per cent. Only Qwen3-ASR 1.7B did better in that comparison. Reson8 ran this test itself; some of the figures for other models come from their makers, who tested with different settings.
For languages with little recorded speech available, such as Maltese and Basque, the model also does much better than Whisper, Reson8 says: an average of 17.6 per cent errors across ten such languages, against 39.8 per cent. On a single powerful Nvidia GPU processing many recordings at once, Rumble-2-Flash transcribed about an hour of audio in half a second on the larger test sets.
Reson8 is the trading name of Resonate Labs B.V. and employs ten people in Amsterdam. Raoul Ritter leads the company and founded it with Thomas Kluiters, who leads AI development, and Jarno Verhagen. In March 2026 Reson8 raised 5 million euros from investors, led by Balderton Capital and with NP-Hard Ventures.
What does this mean for the road to the singularity?
Turning speech into text is already ordinary: phones and meeting software do it every day, usually through a service in the cloud. What is new is a small model with open weights that, according to its maker, can compete with models dozens of times larger. An organisation can run a model of this size on its own computer. A hospital, local council or law firm can then transcribe conversations without the recordings leaving the building. Paid services then have to lower their prices or offer something the free model cannot, such as telling speakers apart or transcribing live.
If that trend continues, speaking becomes an ordinary way to operate software, also in smaller languages. The big leap for Maltese and Basque shows that speakers of such languages in particular stand to gain. A mechanic with dirty hands, someone with poor eyesight or an older person who finds typing difficult can then simply talk to an AI assistant in their own language. A transcript still contains errors, so for a medical or legal decision someone must be able to check what was really said. The people who build these systems can make sure of that by keeping the recording with the text, so that speech becomes a reliable way into AI for everyone.
What is the technological singularity? Read the explanation →