I’m finally ready to release my neural net based lyrics parody generation system…
Introducing: Weird A.I. Yankovic!
Runs on Google Colab: colab.research.google.com/drive/12g07FS2…
Introducing: Weird A.I. Yankovic!
Runs on Google Colab: colab.research.google.com/drive/12g07FS2…
You can provide the rhyme scheme and syllables per line for an existing song, and it will write new lyrics to match.
In the true spirit of parody, here is a Michael Jackson song (“Beat It”) rewritten to be about food.
Then you can sing the song yourself to the horror of others
In the true spirit of parody, here is a Michael Jackson song (“Beat It”) rewritten to be about food.
Then you can sing the song yourself to the horror of others
With a little bit of extra work (you provide a mp3 or mp4), the system will produce a karaoke video to make it easier to sing along
Importantly, we did not do any training. Weird A.I. Yankovic uses GPT-2 and XLNET off-the-shelf.
That means that if you want to use it, you don’t have to find and prepare a specialized dataset to try a particular song, or worry about it not learning.
That means that if you want to use it, you don’t have to find and prepare a specialized dataset to try a particular song, or worry about it not learning.
How does it work? There were 2 parts to this project:
1) writing a near-rhyme detector. Turns out there is no set definition for near-rhymes. Work by @gh_marjan was instrumental
2) Controlling neural language models.
1) writing a near-rhyme detector. Turns out there is no set definition for near-rhymes. Work by @gh_marjan was instrumental
2) Controlling neural language models.
It would be pretty difficult to train a neural language model to meet a given number of syllables per line and also match a rhyme scheme.
But I could pick the rhyme for the next line. The problem then became how to continue from the previous line but end in a chosen word...
But I could pick the rhyme for the next line. The problem then became how to continue from the previous line but end in a chosen word...
XLNet can be made to generate the missing middle of a text span, attending to text before and text after.
And it also has to generate a line with a given number of syllables. Well that just requires a lot of sampling...
And it also has to generate a line with a given number of syllables. Well that just requires a lot of sampling...
The dirty secret of neural language models is that the model itself just produces a distribution over words. The “generator”, which is not part of the model is /code/ that wraps the model and samples from the distribution in some particular way...
Most common are greedy sampling and sampling by logit weight to get beams...
I wrote a sampler that combines logit sampling (beams), heuristics, & custom scoring functions.
Most folks focus on the encoder & decoder that make up the model but forget the importance of the sampler
I wrote a sampler that combines logit sampling (beams), heuristics, & custom scoring functions.
Most folks focus on the encoder & decoder that make up the model but forget the importance of the sampler
First try rewriting Hamilton lyrics came out quite well
