FA+
Shop
5584
Views
Views
79
Favorites
Favorites
Category
Music / Baby fur
Species Pony
Size 120 x 91
File Size 6.2 MB
Report this content
★
More from MillenniumF
Rainbow's Bedtime Problem - Audio Book
Rainbow Dash has to stay over at Twilight's house, but she gets more than she bargained for when Twilight makes her wear something to keep visits to the laundry room minimal.
So a few months back I was informed of this new thing called Fifteen.AI, a website where you can use a neural network to generate speech from text that sounds very natural. I got inspired to create a short audio story, and might do more in the future if you guys like this one. I hope you enjoy it!
Cover art by "mlpcutepic"
So a few months back I was informed of this new thing called Fifteen.AI, a website where you can use a neural network to generate speech from text that sounds very natural. I got inspired to create a short audio story, and might do more in the future if you guys like this one. I hope you enjoy it!
Cover art by "mlpcutepic"
Category Music / Baby fur
Species Pony
Size 120 x 91px
File Size 6.2 MB
You might be interested to know that Fifteen.ai was funded in 2017 by MIT's Computer Science and Artificial Intelligence Laboratory, which is when the neural network began to learn how to talk correctly. It was then fed an hour or so of voice samples from Twilight Sparkle at first in order to generate her voice and can now say any phrase, not just lines from the show (notice how many words in my work have no origin from the show itself). This represents a significant advance in machine learning, because voice synthesis normally takes hours upon hours of recorded dialogue to learn how to speak, and also required a monotone voice to establish a baseline. They would use audio books to get the voice samples, as you could get tons of dialogue as well as an established script to teach the neural network how to do it. Fifteen.ai is able to do the same thing with less than an hour of dialogue, which is mind-blowing!
The guy who created it put in lines from the show for it to learn, then inputted a transcription of that line so that the neural network had a baseline for what each syllable is in text form. Then it simply ran hundreds of permutations, with the guy who made it giving it a pass or fail on how it pronounced them. Once it learned how to do lines from the show, it was then able to learn other words. I'm not sure exactly how it works, but the neural network is pretty great at generating dialogue.
I don't think I'll be continuing this story as it already has a beginning, middle, and end. But I'll probably do more audio books once the new version of Fifteen.AI has been sorted out. I reads the lines WAY too quickly now, so it doesn't sound as natural despite not having a robotic tone anymore.
Comments