LaunchNovember 3, 2026 · 6 min read

Nobody likes waiting for their own words

Meet Rumble-2-Flash: our smallest, fastest, and most affordable speech-to-text model yet. It understands 40+ languages, starting at €0.01 per hour of audio, and it’s live today.

Try it in the playground

Live now in the API

In ASR, speed usually comes at the cost of accuracy. Small models are cheap and fast, and their WER shows it.

We wanted to see how small a model could be without the accuracy falling off. The target was a model cheap enough to run on everything, from voice notes to full call archives, while still holding up on complex, noisy audio.

What you get

​We kept Rumble-2-Flash lean on purpose. Telling speakers apart and transcribing live conversations will be part of a bigger model. More on that below.

Fast

An hour of audio in seconds, not minutes.

Affordable

Starting at €0.01 per hour, pay as you go. No contract. A thousand hours of audio only costs €10 to transcribe.

40+ languages

One model, from Dutch and Polish to Romanian and Basque. Check out the full list.

Any recording

Calls, lectures, voice notes, podcasts.

What all this means in practice

A student in Porto

gets her two-hour lecture back as text before she’s finished packing her bag.

A two-person startup in Kraków

adds voice notes in 12 languages without a second thought about the bill. A thousand hours a month comes to €10.

A podcaster in Cardiff

gets Welsh episode notes seconds after she uploads the episode.

Small model, big ears

Fast is easy if you don’t mind getting words wrong. So we put Rumble-2-Flash through the same public tests everyone in our field uses, and we’re sharing all of it.

The main challenge: real phone calls

Transcribing clean audio is easy, phone calls are not. The audio is squashed, noisy, with fast-talking people on opposite sides of the globe. So we ran Rumble-2-Flash on AppTek’s call-centre test: 302 hours of helpline calls in 14 English accents, scored with AppTek’s own script.

It got about 92% of the words right. That’s on par with NVIDIA’s newest speech model, which is 60 times bigger, and ahead of Mistral, IBM, Cartesia, Deepgram, ElevenLabs and AssemblyAI. One model does better: Alibaba’s Qwen3-ASR 1.7B, more than three times our size.

Real phone calls, 14 English accents
Words missed in every 100. Shorter is better. AppTek call-centre test, 302 hours.02468101214Qwen3-ASR 1.7BNemotron 3.5 Transcribe (30B)Rumble-2-Flash (0.5B)Ink-2Qwen3-ASR 0.6BVoxtral Small Transcribe 3.0FluxGranite Speech 5.0 (2B)Scribe v2 RealtimeU3.5 Pro RealtimeAlibabaNVIDIAReson8CartesiaAlibabaMistralDeepgramIBMElevenLabsAssemblyAI7.47.787.8688.79.121011.071213We ran it, official AppTek scorerFrom the AppTek paper (accent average)Live (streaming), from Cartesia’s Ink-2 post, roundedDifferent setups, so read the outlined bars with care.Full method on our benchmark page.

Percentage of words missed on the AppTek call-centre test. Solid bars: tested by Reson8. Outlined bars: reported by others, with different setups.

The Hugging Face Open ASR Leaderboard

At 0.5 billion parameters, Rumble-2-Flash is one of the smallest models on the Hugging Face Open ASR leaderboard. Across the seven languages tested within the dataset, its accuracy is on par with Meta’s biggest speech model (omniASR-LLM 7B), which is 15 times its size, and ahead of OpenAI’s Whisper.

The best score for its size
18 of 20 models, seven languages (two above 12% WER not shown). Average WER against model size. Higher up is lower WER. 4% 6% 8% 10% 12% 0.5B 1B 2B 5B 10B 20B 40B Model size, billions of parameters (log scale) Word error rate (WER) Rumble-2-Flash 0.5B · 7.09% WER Qwen3-ASR 1.7BAlibaba · 2.04B Nemotron 3.5 TranscribeNVIDIA · 30B Voxtral SmallMistral · 24B omniASR-LLM 7BMeta · 7.8B Whisper large-v3OpenAI · 1.55B Best WER for each size Other models Average over German, French, Italian, Spanish, Portuguese, Hindi and Dutch, Hugging Face Open ASR multilingual test sets.

Average word error rate across German, French, Italian, Spanish, Portuguese, Hindi, and Dutch on the Hugging Face Open ASR leaderboard. Model size on a log scale.

When averaged across Open ASR leaderboard’s English datasets, Rumble-2-Flash lands at about 5% WER, and around 2% on clean read speech. The spread matters more than the average, so here it is per dataset.

DatasetAudioWER
LibriSpeech test-cleanAudiobooks, quiet room2%
VoxPopuliEuropean Parliament speeches2%
SPGISpeechEarnings calls, numbers-heavy3%
Earnings-22Earnings calls, global accents6%
GigaSpeechPodcasts and YouTube8%
AMIMeetings with overlapping speakers9%

Much more than just English

The bigger difference shows up in languages with less training data. Large models are mostly trained on English and a handful of major languages, and their error rates climb quickly outside that group.
‍
We compared Rumble-2-Flash with OpenAI’s Whisper Large V3, a model three times its size, on ten of those languages.
‍
The chart below shows ten of those languages. Rumble-2-Flash averages 17.6% WER across them, against 39.8% for Whisper Large V3, a model three times its size: less than half the errors. The gap widens exactly where other models struggle most. In Maltese it’s 13% against 72%, in Basque 13% against 42%, and in Bangla 26% against 74%.

The languages big models skip
Word error rate (WER), lower is better. Rumble-2-Flash (0.5B) next to OpenAI’s Whisper Large V3 (1.55B). Rumble-2-Flash, 0.5B Whisper Large V3, 1.55B Maltese 13% 72% Bangla 26% 74% Albanian 22% 52% Basque 13% 42% Afrikaans 20% 33% Lithuanian 18% 28% Icelandic 26% 36% Latvian 12% 20% Macedonian 12% 18% Croatian 10% 12% 0% 10% 20% 30% 40% 50% 60% 70% 80% 10 languages, 17 dataset/language groups, 90.8 hours. Mean WER: Rumble-2-Flash 17.60%, Whisper Large V3 39.82%. Full results on our benchmark page.

What you need to know about benchmarking

Speech tests compare what a model writes with what a person typed out, but people make mistakes too. Sometimes they mishear. Sometimes they write “twenty” where the model wrote “20”, and the test counts that as wrong.

When we fixed the human mistakes in one well-known test, our misses dropped from about 6% to under 3%. And where the original typist got it wrong, Rumble-2-Flash usually wrote what the speaker really said correctly.

We’re not telling you this to make our numbers look better. We just think it’s how anyone’s scores should be read, ours included.

If you are evaluating ASR for your own product, test on your own audio.

Read our benchmarking guide

For the rooms where every word matters

We built Rumble-2-Flash for everyday audio, so anyone can use it. It handles voice notes, interviews, and one-on-one calls well, and even in a busy meeting or a noisy room it gets around 9 in 10 words right.

That’s plenty for a voice note. In a doctor’s appointment, a board meeting or a voice agent serving millions of callers, one wrong word in that last 1 in 10 can change a diagnosis or a decision.

That’s what Rumble-2 is for. It’s the most accurate model we’ve trained so far, and it adds speaker diarization, end-of-turn detection, automatic language detection, and custom models.

Who’s behind this?

We’re ten people in Amsterdam. Builders, engineers, and, technically members of staff.

A note from our CEO:

We started Reson8 because voice is the way humans have always connected. Making something this small and this fast is just the beginning. If you build something with it, or it gets something wrong, tell me directly at rr@reson8.dev.
I read every message.

Raoul Ritter, CEO, Reson8