AI Aulis

AI NEWS IN ENGLISH · 19 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

38 stories · next issue tomorrow morning · How the paper is made →

← Issue 19 November 2026Paper

Reson8: 0.5-billion-parameter speech model matches NVIDIA’s 30-billion-parameter model in phone-call test

AI-generated contentWritten by AI

huggingface.co, siliconangle.com · 2 sources · Sources published: 18 November 2026

Amsterdam-based Reson8 has released the weights of its Rumble-2-Flash speech recognition model under the Apache 2.0 licence. The company said the 0.5-billion-parameter model recognises speech in more than 40 languages and transcribes English phone calls almost as accurately as NVIDIA’s 30-billion-parameter model.

Reson8 (Resonate Labs B.V.) announced the release on 18 November in a Hugging Face blog post, and SiliconANGLE reported the company’s figures the same day. The model is designed to transcribe pre-recorded audio. All accuracy and speed figures come from the company.

Reson8 trained the small model by distilling it from its larger Rumble-2 model. During training, Rumble-2-Flash transcribes the audio itself and the larger model scores the transcript token by token. This gives the small model feedback on its own mistakes. Because the model uses the Qwen3-ASR architecture, it runs in the Transformers and vLLM libraries without custom code.

The table shows word error rate (WER), the share of words transcribed incorrectly, on AppTek’s call-centre benchmark: 302 hours of English helpline calls in 14 accents.

Model Parameters WER
Qwen3-ASR 1.7B 1.7bn 7.4%
NVIDIA Nemotron 3.5 Transcribe 30bn 7.78%
Rumble-2-Flash 0.5bn 7.86%
Qwen3-ASR 0.6B 0.6bn 8.7%

Source: Reson8 blog, scored with AppTek’s own script.

Reson8 said the model is on par with Nemotron; the gap is 0.08 percentage points. Qwen3-ASR 1.7B, from the same model family, does better in the comparison, however. The benchmark is in English, so it says nothing about calls in other languages.

The sources give different multilingual figures. The blog said the model averages a 6.96 per cent error rate across nine languages on the Open ASR Leaderboard multilingual track, and 6.6 per cent WER on the English track. SiliconANGLE reported that the company’s own runs produced an average of 7.09 per cent across seven languages and about 5 per cent in English. No reason is given for the difference. SiliconANGLE also reported that Reson8 puts the model’s mean word error rate at 17.6 per cent across ten lower-resource languages, such as Maltese and Basque. OpenAI’s Whisper Large V3 scored 39.82 per cent on the same test, the company said.

The blog said the model transcribes more than 7,000 times faster than real time, or an hour of audio in about half a second. Speed depends, however, on how full the processing batches are. The figure applies to the large datasets on the Open ASR Leaderboard, where batches stay full.

The weights are available to download from Hugging Face (Reson8/Rumble-2-Flash). SiliconANGLE reported that the model can also be used through Reson8’s API, with prices starting at €0.01 per audio hour. Only Rumble-2 accepts a custom vocabulary without fine-tuning. Neither source lists all of the more than 40 languages, so users need to check Finnish support themselves.

Report an error in this story