IT ALMOST feels like we should start a “what Janelle did next” section on the INQUIRER, such is our love for the quirky things she can do with AI.
But what sounds to choose? If only the BBC had released 16,000 .wav files of sound effects into the public domain recently.
Oh, wait, that’s right, they totally did.
The data set in question (ie the sound effects) were fed into a textgenrnn neural network and the network came up with a pithy name.
And occasionally you can see where it’s coming from.
Take this one. To the BBC, it’s a cuckoo clock. To a neural network, it’s “German household operating”.
Or this one – officially, it’s called “Wind in trees” Except now, it’s called “Agitated door cat, interion, chickens – 1972”
It may not be right, but it’s a lot more fun.
As the sounds get more esoteric, so do the describers.
This is what “Small man continuous large poop” sounds like. Except it’s actually meant to be “Pigeons cooing. (Close perspective recording.)”
Of course, nobody expects a neural network to get the answers right without lots more training, but these names aren’t random either – they are offering us a glimpse into the way the neural network is perceiving input, which is kind of fascinating.
In the same way that it’s fascinating for a Siamese Cat to be identified as “Dinghy passes away”.
Not sure what to do with that.
Our final selection is ‘Baby girl, restless in cot, early morning, 9 months old – 1984′. Only to a neural network’s ‘ear’ it’s… erm… “Peacock butter, with background clock children. “
As ever, there are loads more here – including a higher level of mild smut than previous experiments, and as ever, the really weird and dirty stuff is saved for subscribers to Janelle Shane’s newsletter and who are we to spoiler that? μ
Source : Inquirer