Pop*

Pop* (pronounced Pop-Star) is a system developed by Paul Bodily starting Winter 2016 to automate the composition of pop songs. A fundamental element of the project is the automation of annotating tablature data as found in open-source databases such as ultimate-guitar.com. This venture is titled Tab-Complete. The purpose of this wiki is to facilitate documentation of the research endeavors for both the Pop* system and the Tab-Complete dataset.

• Tab-Complete - Automated completion and annotation of pop music tabs
• Pop* Log
• Compositions
• Audio Rendering
• Log Entries
• Ben's Log
• Down by the Bay

• Sequence alignment methods for identifying structure in music
  • Pairwise MSA for identifying consensus of several lyric sources
  • Lyrical sequence alignment for identifying musical content of tab sources
  • Aligning pairwise alignments of lyric lines for chorus identification (alignment of alignments)
  • Aligning pairwise alignments of chord lines for verse identification (alignment of alignments)
  • Phoneme sequence alignment for rhyme scheme detection
  • Pairwise alignment of intra-segmental chord lines for substructure identification
• How do we get complete tabs? Compare MSA with pairwise Smith-Waterman techniques.
• Using rhymes and a distribution metric to pull structure out of unstructured tabs
• Comparison of key normalization techniques for chord sequences
• Comparison of melody/lyric synchronizers

• Conference on Computer Simulation of Musical Creativity (CSMC) - 17 – 19 June 2016

• The number of chords per line doesn't equal the number of measures per line
• All lines in a segment will have the same number of measures per line.
• All MMs are single-order right now
• Lyric transition models are based on word→word transition frequencies, not syllables→syllable transition frequencies (which we may want to try)
• We select the number of words (not syllables) per line, and find a way to make it work syllable-wise

• Make sure we're excluding tab info in opening lines, like in https://tabs.ultimate-guitar.com/e/eagles/its_your_world_now_ukulele_crd.htm
• lyricsnet isn't scraping the beatles
• Include figures in manuscript
• Get some gold-standard data to validate with and insert results into manuscript
• Get a way for Ben to access verses/choruses on demand, etc.
• Need better segmentation of songs to avoid a bunch of 1-line segments
• Need better rhyme scheme analysis to get good rhyme constraints
• Chords are being grouped only by root and minorness. Needs to be expanded to get more variety.
• Normalization of key signatures needs to happen
• Need to check if rhyme constraints are working properly
• Need to add constraints
  • on chorus to indicate which lines are fixed and which aren't (lower priority)
  • for subrhyme schemes
  • repetitive subsequences of chords
  • repetitive subsequences of lyrics
• Start drafting paper on data process and alignment
• Chorus must also match chords
• Create gold standard labeled dataset to test various methods. Label
  • Key
  • Rhyme scheme
  • Internal rhymes
  • Segment structure
• Normalize keys
  • Use neural network with one input feature for
    • each of 24 major/minor chords (double 0 to 1 representing frequency of chord)
    • first chord
    • last chord
  • and for output one node for each possible key
• Use tab contents rather than lyric contents in finalized tab (use lyrics just for identification of song body and completeness)
• When eliciting rhyme scheme, currently it's getting all pronunciations of the line and then taking the last few syllables of each (highly redundant). Fix it.
• Compare different rhyme scoring algorithms:
  • Pat Patterson's rules
  • Hirjee Matrix simple scoring
  • Pat Patterson's rules w/ Hirjee Matrix
  • Alignment
  • w/ w/o considering penultimate syllable
  • Normalizing the Hirjee Matrix
  • Penalize for distance
• Start parsing the scraped data
• Get rid of blocks in tab with no chords && no matching lyrics
• Explicitly set costs in Aligner before it is called.
• Have TabComplete print out a tab-delimited file with columns representing value indicative of the quality at each step (don't filter initially, choose later)
  • May want to make this a function of how many sequences are being aligned?
• Do thorough manual verification of pipeline up through tab parser
  • Check how complete the lyric consensuses are
  • Check that all the fields are being correctly populated
  • Check how many lyric sheets per song are being aligned
• Create or find gold standard dataset

• There are a number of known issues with the MSA consensus calling. It may be better to just use the text from the tab, rather than whatever mish-mash the consensus pulls out (e.g., mindeah).
  • Right now any alphabetic character will taken precedence over a non-alphabetic character if there's a tie
  • If there are three different sequences with three different characters, one will get picked at random or may not get picked at all, depending on where gaps are inserted.
  • So much depends on the scoring costs
• Find threshold for rhyming vs no rhyming
• Try running on just good tabbers tabs or on Pro tabs (requires subscription?) or those with videos?
• Implement and give graduated scoring for relative major chord matching in chord alignment
• Find best cost matrices for MSA (i.e., match cost, mismatch cost, gap open/extend costs)
• For speed,
  • consider nixing the all v all pairwise alignments and just do pairwise MSA in random order
  • consider not using a map for the storage of MSA aligned sequences

• Use conditional probability distributions to model number of chords per line, and constraints relating to number of chords per line and repetition of chords in a line
  • We have a distribution of rhyme schemes conditioned on SegmentType
  • We need a distribution of subrhyme schemes conditioned on rhyme schemes (could also be conditioned on SegType)
  • We need a distribution of the number of chords per line conditioned on rhyme schemes (and possibly conditioned on SegType)
  • We need a distribution of repetitive subsequences of chords per line conditioned on rhyme scheme and SegType
  • We need a distribution of the number of words per line conditioned on subrhyme scheme
  • We need a distribution of repetitive subsequences of lyrics per line conditioned on rhyme scheme and SegType
  • We need a distribution of the variation in words per line between paired lines (as per rhyme scheme)
  • We need a distribution of chord transitions, including intersegmental chord transitions, starting chords, ending chords, etc.
• Instead of keeping just counts, keep pointers to actual songs to be able to later reference source of decision
• Test: Rhyme Structure sampler with actual constraints
• Filter explicit songs -
• Complete tab alignment with assumption that chord indices (within line) are not correct and that blocks are not maintained
• Find matching sheets to do MSA - we will start by only aligning those with matching names
• Find choruses
• Words not in the CMU dictionary?
  • Integrate G2PConverter into phoneticizer
• Fix: Right now a one-line bridge between verses is screwing up distributions for lines per bridge, etc. Solution: treat such as interlude
• Fix: When getting phenotypes per line, need to let phoneticizer (w/ rhyme stop words) decide what to get phonemes for (right now just gets last two words, even if they’re stop words)
• Account for different pronunciations in the CMU dictionary
• Discovered the refactoring takes the same time for both string alignment and MSA. I'm a little concerned the MSA is going to take a long time to complete.
• Test refactored code to see if alignments are the same before and after refactoring and the change in speed from generalizing
  • Compare the two MSA alignments algorithms on simple inputs, comparing their scoring matrices and figuring out why they compute differently, specifically consider subsequence from a real case.
  • I discovered I'd never updated the fix for computing left costs for the old aligner AND simultaneously stumbled on a bug in the consensus caller that overestimated the count for a character that appeared in both upper and lower cases.
• Verified that the refactored alignment code computes the same alignment for strings
• Refactor sequence alignment code to easily allow alignment of non-char sequences (e.g., phonemes, chords)
• Find out how unbanded and banded with minPercOverlap of 1.0 are different and fix it.
• Can we grease up the banded?
• n-dimensional MSA? or Pairwise?
• Computer?
• Implemented and compared alignment algorithms for aligning lyrics and tabs
• Fix issues with scrapers:
  • UG: raw_tab,url, title, difficulty, key, provider, contributor, type all look good
  • eChords: raw_tab,url, title, difficulty, key, provider, contributor, type all look good
  • SongLyrics: url, title, provider, artist all look good
  • Metrolyrics: url, title, provider, artist all look good
  • LyricsNet: url, title, provider, artist all look good
• Map role values to actual roles
• Parse untagged chords in tabs
  • Legitimate Chords:
    • “Intro:G” or “Bm…” or “C|” or “Gmaj7 –” or “~D7” (1,2, 3,4)
    • “C+” or “Bb5+”
    • “Em-Dm-C” (1)
    • “Ebsus” (1)
    • “Dm/C” or “F6/A” (1)
    • “A4” or “Am2” (1)
    • “D9sus4”, “Bm7b5”, “F#m7b5”,“Am7add11” (1,2)
    • “Dm13” (1)
  • Inferable Chords:
    • “/F” (1)
    • “-7” (from “C -7”, i.e., space) (1)
    • “Em7-5” where Em7 is labeled
    • “Δ7” or “º7” (1,2)
    • “Do” ( 1)
    • “Am^D7” (1)
  • Repeat Information:
    • “x4” or “repeat” or “(2x)” (1,2)
    • “||: :||” (1)
  • Salvageable?
    • chords both in and before line
    • “7B” (1)
    • whole line of unmarked chords
  • Removable elements:
    • “(hold)” or anything in parens? (1)
  • Things to check:
    • Impact on imbedded? (1)
    • Still seeing “=”?
    • Still seeing “de de da__ da da__ C” (1)?
• Replace lyrics with more than three consecutive identical letters with just two of them ( solution)
• Fix songlyrics.com scraper to not just get Billy Joel songs
  • Added “0-9” to the two LinkExtractors…
  • Changed start_urls to http://www.songlyrics.com/
• Fix metrolyrics.com scraper to also get songs and artists with numbers in their name
  • Added “0-9” to the two LinkExtractors…
• Check tab scrapers and lyrics.net to see if they're getting all the artists
  • Looks okay

• Switch the chord/lyric tab parser to Java
• Switch the lowerlevel loader/cleaner to Java
  • Fix missing artist in E-chords tabs
  • Remove extra column for lyrics
  • Fix difficulty and contributor parsing for UG

** Answer questions:

**

• Are there regularities in the tabs titles or artists that can be removed?
  • Removed all UG Guitar Pro Tabs
  • Removed type from UG song titles
  • Removed (),[] everywhere they appear in titles or artist names
  • Used lowercase, alphanumeric-only keys for artist and song names for indexing
  • Removed “ Tab” from UG artist name

**

• Are there regularities in the tabs content that can be removed?

**

• Are there regularities in the lyric titles or artists that can be removed?
  • Removed (),[] everywhere they appear in titles or artist names

**

• Are there regularities in the lyric content that can be removed?
  • For lyricsnet, the content always starts with the tag “<pre id=“lyric-body-text” class=“lyric-body” dir=“ltr” data-lang=“en”>” and ends with the tag “</pre>”
  • For songlyrics, the content almost always starts with the tag “<p id=“songLyricsDiv” class=“songLyricsV14 iComment-text”>” and end with “</p>”(1186539/1186581), otherwise it has no lyrics at all.
  • For metrolyrics, the content always starts with the tag “<div id=“lyrics-body-text”>” and end with the tag “</div>”(815749/815749)
  • Turns out all of these are consumed in the JSoup parsing