live-rap began as a Bcgamejam2020 idea that wasn't finished (or really touched) during the jam.
The theme was "Prevail" and I brainstormed on what idea wouldn't possibly be done by anyone else - and boom, I decided to look up the anagrams for the word!
So I came upon "live rap", and this stuck to me.
The purpose of the program is to "create" new rap lyrics based off of a large set of training lyrics.
It scrapes lyrics of raps that qualify as English lyrics and creates a word2vec model based on the words belonging to the lyrics.
Then it trains the word tokens by:
- Replacing contractions with incorrect syllable count detection
- Replacing numbers with words
- Removing invalid tokens, usually punctuation
- Removing profanity
Once the training is complete, given a song, live-rap TM will create lyrics of the correct syllable count and rhyme that you can now sing during karaoke (causing confusion and awe to everyone else)!
First install the dependencies via: pip install --user --requirement requirements.txt
then run: python liverap.py
Task 1: Make rap lyrics from other lyrics
- Download a bunch of popular rap song lyrics, and train a word2vec (CBOW or skip gram) model based on that corpus
- Take a new song lyric (and the background track) and keep count of line number and the # of syllables in each line
- Create the lyrics so that each line makes sense (randomly choose a word? and make a sentence?)
Task 2: Rap lyrics from topic
- Use a corpus of words that's public first to train a model OR use a trained model (idk if there is in word2vec)
- Create lyrics that make sense FIRSTLY within that sentence given a topic 2.1 Then perhaps expand so that the entire stanzas make sense together
OPTIONAL: TODO: Make the lyrics rhyme, so make sure the last word of the sentence rhymes and then work backwards
- Whenever building the model, also create a dictionary of (word in model) to (words that rhyme with that word)
- Use this dictionary to choose words so that they rhyme
- Also when creating lyrics, make sure the lines that rhyme in the OG lyrics also rhyme in the new lyrics
- make a grouping of rhymes dict, which all rhyme to the key. The values is a list of the line #s
- Now build the sentence from end to front
DONE: Use a diff tokenizer that doesn't split single quotes (TreeTokenizer) https://stackoverflow.com/questions/34714162/preventing-splitting-at-apostrophies-when-tokenizing-words-using-nltk?lq=1
DONE: Make a rap lyric scraper quickly to speed up process For now just manually
DONE: Clean the data more (ex. remove some numbers, brackets, punctuation)
DONE: replace numbers - https://stackoverflow.com/questions/40040177/search-and-replace-numbers-with-words-in-file
DONE: remove non english
DONE: punctuation
DONE: omit some contraction replacements that dont impact syllables (ex. ain't)
TODO: Scrape more songs to have larger model
DONE: Perhaps censor profanity?
TODO: Prettier read me with screenshots (of non profane lyrics!)