Nobody likes waiting for their own words.
Meet rumble-2-flash. Our fastest and smallest, at the lowest price yet: 80+ languages, 0.015€ per hour, and live today.
Free until 10th of November 2026
Most speech models get better by getting bigger. More parameters, bigger GPUs, bigger bills. We went the other way.
rumble-2-Flash has 0.5 billion parameters. The models next to it on the Hugging Face multilingual leaderboard range from 2 to 24 billion. It still finishes fourth of ten, just 0.7 points behind Voxtral Small, a model 48 times its size.
Average word error rate across German, French, Italian, Spanish, Portuguese, Hindi and Dutch on the Hugging Face multilingual leaderboard. Model size on a log scale.
Small is also why it’s fast, and why it’s cheap. It runs 1,324× faster than real time, so an hour of audio takes about three seconds. And it costs €0.015 per hour.
Try it before you read another word.
Drop in a recording: a voice note, a lecture, a call. The street name every app gets wrong, or your grandmother’s favourite saying. Then see how fast the text comes back.
First 7 days on us. Then come back. We’ll wait.
So what do you
actually get?
We kept rumble-2-flash lean on purpose. Those come in something bigger. More on that below.
0.0015/h
Pay-as-you-go. No contract, no minimum. A price that stops being a decision.
80+
languages in one model, from Dutch and Polish to Welsh and Basque.
Any recording
Calls, lectures, voice notes, podcasts. Send it a file, get text back.
Europe
Hosted in Europe. Every company we use to run it is in our trust center.
Small, and still sharp.
We put rumble-2-flash through the same public tests the whole industry uses, and we’re publishing every result, including the ones we’re not proud of yet.
1.6%
Clear speech
LibriSpeech clean. Fewer than 2 slips per 100 words.
5.9%
English, 11 test sets
Meetings, calls, podcasts, noisy rooms. About 94 in 100 words exactly right.
7.9%
14 English accents
AppTek, 1.2 million words, from Glasgow to Singapore.
8.7%
30 languages
FLEURS average, the standard multilingual test.
Checked by someone else.
These numbers come from the Hugging Face Open ASR benchmarks, so you don’t have to take our word for it. RTFx is speed: how many times faster than real time the model runs. 1,000× means an hour of audio in under four seconds.
How it stacks up.
On the Hugging Face multilingual leaderboard (German, French, Italian, Spanish, Portuguese, Hindi, Dutch), serval-2-flash would sit fourth among these models, and it is by far the smallest: 0.5B parameters, four times smaller than the next-smallest model and 48 times smaller than Voxtral Small. It finishes ahead of Whisper large-v3 and Voxtral Mini, and second only to Qwen3-ASR on Hindi.
The way people actually talk.
Accents are where speech recognition usually breaks. General American comes in at 4.6%, Irish, Indian and Southern US all stay under 7.5%, and Welsh accents sit at 9.3%.
Scottish and Singaporean English are the hardest so far, around 11.5%. We’re working on those.
Beyond English.
On FLEURS we measured 30 languages, with an average of 8.7%. Italian, Spanish and English lead. Arabic, Persian and Hungarian are our weakest: if you’re building in one of those, test it on your own audio first. We’d rather you hear that from us.
Japanese, Korean, Thai and Chinese don’t put spaces between words, so we count errors per character there instead of per word.
Every number, and exactly how we measured it, is on our benchmark page. Where we’re not the best, it says so there too.
The best benchmark is your own recording.
What that means for you.
A student in Porto gets her two-hour lecture back as text before she’s finished packing her bag. [real timing]
A two-person startup in Kraków adds voice notes in 12 languages without a second thought about the bill. [real cost]
A podcaster in Cardiff gets Welsh episode notes minutes after she uploads the episode. [real example]
Price shouldn’t decide who gets heard.
A farmer in Galicia, a nurse in Riga and a founder in Tallinn deserve voice technology that works for them as well as it does for a podcaster in California.
Right now, the best voice models are big and expensive, and they speak English first. So the side project never gets built. The app never launches in Latvian.
We’d rather you never have to do that maths. Build the idea. Spend less. Scale faster. We’ll keep the bills as low as possible. It’s 0.015€ per hour.