Murmel v3: faster, more efficient and better
Murmel v3 is here. A new model, built from the ground up for Dutch. It understands better, is a lot faster and now knows who is speaking. There is nothing to do: every new recording goes through the new model automatically.
Much of what is in this version comes straight from your feedback. Over the past months municipalities, journalists, researchers and companies told me where things had to get better. Every bit of that feedback was taken along.
Built for Dutch, not added on
Most speech models are made for English. Dutch is in there, but as one of many languages. You notice it in the mistakes: names, slips of the tongue, someone mumbling a little, a debate where people talk over each other.
Murmel v3 has heard Dutch from day one, nothing but real speech the way it sounds in the Netherlands. As a result it makes 38% fewer errors in Dutch than the previous version, and precisely on the audio that gets hard the difference is largest.
An hour of audio in under a minute
The new model is smaller and more cleverly built, and therefore much faster than the previous version: up to 85 times faster than real time. An hour of audio is done in under a minute, even with speaker recognition on. Long recordings of a full meeting day are no longer a problem.
Faster is also leaner: less power and less compute per hour of audio. That gain goes to more minutes and new features, not to a higher bill.
It writes down what was said
Murmel v3 writes down what it hears, the way it was said. A "zeg maar", a repetition, "'t" and "z'n": it is all there. Numbers become digits, capitals and punctuation it adds itself. For minutes, subtitles and research that is exactly what you want: first what was said, then what you do with it.
And it makes nothing up. On silence, music or a jingle the text stays empty, where older models would sometimes invent a sentence.
A name the model has never heard is written the way it sounds. That is what the word list is for: put your organisation's names, places and jargon in your profile and Murmel spells them right from then on. In a test with debates and podcasts that removed dozens of name errors per recording.
Who said what
Speaker recognition is now part of the model itself. Turn it on and Murmel gives every speaker their own label, up to twelve speakers per recording, with no extra waiting. In a test with eight known speakers it never mixed two people up. Every word also has its own timestamp: click a word and the audio jumps there, with speakers on as well.
Dutch and English mixed
An English quote in a Dutch interview, a conversation that switches language halfway, an English podcast with Dutch ads: Murmel v3 hears which language is spoken and simply carries on. Nothing to set up front.
You can try Murmel v3 free for two weeks. Start your free trial at murmel.nl.