Week of 2/14/19
(Mostly research)
WORDS AND MELODY ARE INTERTWINED IN PERCEPTION OF SUNG WORDS:
- Purpose: determine whether words and melodies in songs are processed interactively or independently, determine influence of attention on the processing of words and melodies in songs
- Conclusion: lyrics and tunes are intertwined in sung word recognition, variation in melody and/or variations in lyrics elicited a cerebral response
A MELODY-CONDITIONED LYRICS LANGUAGE MODEL:
- Purpose: create a large database of lyric-melody aligned content with syllable-note alignments and word/sentence/paragraph boundaries, in the end to generate a language model that produces lyrics given input melody
- Previously there had not been a sufficient amount of data to analyze the relationship between lyrics and melody
- Writing lyrics from melody requires accounting for word boundaries and rests in melody
- Current computer-aided lyric writing software still requires the user to interpret the source melody and determine constraints such as syllable count and rhyme position
- Goal: create a program that automatically identifies these restraints otherwise identified by the user and creates an independent program to generate coherent lyrics
- Discourse structure of lyrics (sentence/paragraph boundaries) is determined by melody rests and context words
- Lyrical generation requires more than just the analysis of syllable stresses and beats
- Melody-lyric database: dataset of digital music scores, each specifying a melody score augmented with syllable information (word/sentence/paragraph boundaries) for each note, aligned with raw lyrical text files (see Figure 2 from paper below)
- Phenomena discussed in paper:
- Words, sentences, and paragraphs rarely go beyond a long melody rest
- Boundaries of larger components generally align with longer melody rests
- Recurrent Neural Network Language Model: neural network model that remembers all previous inputs
- Score data is useful to analyze relationship between phonological aspects of lyric and melody, but the lyrical text files are needed to analyze the discourse components (need structural information of the lyrics)
- Word segmentation and pronunciation: accomplished with a morphological analyzer
- Alignment of lyrics and score: Needleman-Wunsch algorithm (global alignment of two sequences)
- ****Method for data creation is general enough to be applied to MusicXML (I just need to get the program working, will be focusing on that next week)
- Boundaries of Lyrics: Notable Finds
- Positions of lyrics segment boundaries are biased to melody rest positions (less likely to appear at note positions)
- Probability of boundary occurrence depends on the duration of a rest (see Figure 4 below)
- Short rest= word boundary
- Long rest= block boundary
- Probabilistic approach: modeling lyrics using a rigid set of rules would be extremely difficult; better to use a melody-conditioned language model
- Need to create a language model that generates coherent lyrics with discourse segments that fit a given melody (segment boundaries fit distribution directly above, in terms of rest and boundary alignment and size)
- Melody-conditioned RNNLM: standard RNNLM with featurized input melody
NEURAL NETWORK LANGUAGE MODELING:
To Do:
- Learn more about neural network language modeling & machine learning
- Figure out how to get Matlab working and apply the Japanese article's findings to MusicXML
- 1 sentence demo description