In many cases, Samuel learns to position its virtual mouth as a human would, even though its not explicitly trained to do so.

Samuel is a model I trained to reproduce speech in Pink Trombone, a silly vocal tract simulator. It's more phonetically realistic for vowels than for consonants - for consonants, it never closes the mouth for plosives like k and b. This has to do with the loss landscape.

Try it: Explanation:

Rounded vowels produced by the tip of the tongue instead of the lips you say?

ok how the heck do you know this

a curiosity on a global level + I just happened to be checking something about salishan languages a couple days ago

Whats actually insane here is that you could use this to train deaf people how to speak perfectly without ever hearing themselves. It can literally tell them how their voice production needs to be adjusted to sound correct.

can you train it to beatbox

This would be immensely useful for learning to sing! Looking forward to future iterations.

@avzaagzonunaada very interesting

can help the tone deaf diagnoise why every note sounds the same

For years I was trying to make a language script according to the tongue shape on each letter. because i cant find any explanation on letters design

Wow

👀

Shout out to everyone getting weird looks now making random sounds out loud while watching this

Need this to train me on my trumpet playing!

How fast is it to run? Can it be used on cheap phones to teach people pronounce certain sounds?