Omega Haileyesus

Week of Apr 8 - 12

• Continued implementation of melody generator tool
• Added ability for longer words to get different note types to make it flow more naturally
• Brainstormed outline ideas for collaborative group paper
• Next Steps:
  • More sophisticated library of notes to pull from
  • Get more music theory into the melody generation
  • Time permitting: Look into some deep learning approaches

Week of Apr 1 - 5

• Worked on implementation of melody generator tool
• Status:
  • Given any .txt file, it breaks up the lines into the individual words, checks the ending phoneme and assigns all phonemes the same note.
  • For example:
    • "One fish, two fish. Red fish, blue fish" would be mapped to something like
    • "A B, C B. D B, C, B."
  • Uploaded code to github: https://github.gatech.edu/ohaileyesus3/vip-spring19
• Next steps:
  • implement more complex notes/chords
  • check for more complex rhymes; examine more than just the last phoneme

Week of Mar 25- 29

• Out of town for National Society of Black Engineers' Annual Convention
• 3/4th Presentations; added my slide to the presentation: https://docs.google.com/presentation/d/1TckYH1P9U7jn5_lhVD6MZFOJXVP-r7vR8HpCIljNJBk/edit#slide=id.g553b11f276_1_0

Week of Mar 18- 22

• No meeting; Spring Break

Week of Mar 11 - Mar 15

• Wanted to see if the kanye_network linked in previous weeks was actually detecting number of syllables but it turns out it was just a hyperparameter not actually determined by the network
• After exploring a variety of neural nets and lyric generators to little avail, I realized that given my lack of technical music training, it would be better to back track and gain a better fundamental understanding of the technical side of rhyme scheme, meter, and rhythm
  • https://spinditty.com/industry/How-to-Set-Words-to-Music
  • Rhythm in music refers to the way the beats are arranged, the stresses that give us strong and weak accents
  • break the words up into their constituent syllables
  • when analyzing iambic pentameter, the strong-weak accents are regular so can provide easy basis for creating a melody
  • final stage of the process involves turning your rhythm into a melody
    • Base it on a scale pattern - that could be a major, minor, pentatonic, blues or whole tone scale, going up or going down
• Return back to the stress/rhyme models from Deep-speare: https://people.eng.unimelb.edu.au/tcohn/papers/acl18lau.pdf

Week of Mar 4 - Mar 8

• No meeting; Guthman Competition

Week of Feb 25 - Mar 1

• Continuing goals from last week; no significant updates

Week of Feb 18 - Feb 22

• Took a deeper look into https://github.com/robbiebarrat/rapping-neural-network
  • Tried changing the lyrics on which it trains
• Next week:
  • May try to incorporate Rob's lyric scraping tool to train network above^^^ on other artists
  • First iteration of demo could just be hitting a key at the end of a line or on every rhyme
  • Should shift focus less about lyrics themselves and observe how these repos actually do the analysis

Week of Feb 11 - Feb 15

• Looked into implementing Raplyzer
  • Got it running locally & tried manipulating existing repo of lyrics
    • When I originally ran it, it printed results normally. Then I erased the contents of the .txt file that had the highest rating and it had no effect on results. If I decide to pursue Raplyzer further, I need to take a deep dive into the code to figure out why this happened
• Looked into trying to implement DopeLearning
  • looks like it's powered by DeepBeat: http://deepbeat.org/
    • Doesn't look like its open source, or at least I couldn't find a repo
• Found another rap lyric generator github repo - https://github.com/kjaisingh/rap-lyrics-generator
  • Ran it locally and began training the model
    • Kept interrupting and only message is "Killed: 9"
      • so I ran it on the weinberg-lenovo-1.music.gatech.edu machine
        • MemoryError
• soooo, I found another one: https://github.com/robbiebarrat/rapping-neural-network
  • and this one seems very promising. It currently trains on just kanye's discography, but I ran it in the google colab notebook linked in the github readme and got it to generate a whole verse in a matter of minutes.
  • will spend next week looking into how this one was implemented
• During next week, send Shannon rap songs that are happy, sad or informative
• Potential Demo Idea: Given an existing rap verse, we can auto-detect the flow/bpm/tempo and have shimon play along. (If successful, maybe get this working with a live rap verse that dynamically changes tempo)

Week of Feb 4 - Feb 8

• Raplyzer
  • multi-syllable rhyme schemes: not only the last syllable but multiple syllables that share a vowel sound
  • great open source speech synthesizer, eSpeak, which can be used to obtain a phonetic transcription of the lyrics
  • how it works:
    • For each word, find the longest matching vowel sequence that ends with one of the 15 previous words (doesnt not consider duplicate words or lines)
      • (could we look into a different pool of words other than the previous 15?)
    • Compute the average rhyme length (= Rhyme factor) by averaging the lengths of the longest matching vowel sequences of all words.
  • Battlebot:
    • “spit” any line that comes to your mind and BattleBot will respond with a list of the best rhyming lines
    • page forbidden but the concept seems pretty similar to what we want to do
• DopeLearning: A Computational Approach to Rap Lyrics Generation
  • RankSVM algorithm, and a deep neural network
  • uses same rhyme density measure as raplyzer
  • "use our model to identify the best next line among a set of candidate next lines taken from the lyrics repository"
    • not from the same song and that they do not end with the same words
  • outperforms the top human rappers by 21 % in terms of length and frequency of the rhymes in the produced lyrics
  • (we can look into incorporating emotion when considering next line)
  • (if shimon is simply producing melodies, we may not need lyric generation)