MusicIR

MusicIR is a music information retrieval package for ruby. Its key features are:

• Quantizing MIDI performances ("snapping" them to its best guess at eighth notes, quarter notes, etc)
• Estimating the meter (time signature) of a phrase
• Estimating the harmonic context (the implied chords, the underlying key) of a phrase
• Estimating the phrasing of a piece.

Current limitations:

• Only represents a single stream of notes
• Does not fully handle rests yet. (Some methods assume legato, back-to-back notes.)

Installation

Add this line to your application's Gemfile:

gem 'music_ir'

And then execute:

$ bundle

Or install it yourself as:

$ gem install music_ir

Usage

Here's an example of detecting the a phrase's meter:

require 'music_ir'

include MusicIR

# define some pitch values
PG4  =        67
PA4  =        69
PB4  =        71
PC5  =        72
PD5  =        74

# first phrase of Bach's Minuet in G
notes = []
notes << Note.new(Pitch.new(PD5), Duration.new(2))
notes << Note.new(Pitch.new(PG4), Duration.new(1))
notes << Note.new(Pitch.new(PA4), Duration.new(1))
notes << Note.new(Pitch.new(PB4), Duration.new(1))
notes << Note.new(Pitch.new(PC5), Duration.new(1))
notes << Note.new(Pitch.new(PD5), Duration.new(2))
notes << Note.new(Pitch.new(PG4), Duration.new(2))
notes << Note.new(Pitch.new(PG4), Duration.new(2))
nq = NoteQueue.new(notes)

if nq.meter # because sometimes the meter is ambiguous
  puts nq.meter.val # => {:beats_per_measure=>2, :beat_unit=>4, :subbeats_per_beat=>1}
  puts nq.first.analysis[:beat_position].to_hash.inspect # => {:measure=>0, :beat=>0, :subbeat=>0}
end

Contributing

1. Fork it
2. Create your feature branch (git checkout -b my-new-feature)
3. Commit your changes (git commit -am 'Add some feature')
4. Push to the branch (git push origin my-new-feature)
5. Create new Pull Request

More Details on Implementation

Meter Estimation

MusicIR estimates a NoteQueue's meter by performing an autocorrelation on the phrase. The resulting vector is compared against reference vectors for every supported meter. (These were generated by taking the mean of vectors with known meters, including both Bach chorale melodies and the melodies of a set of folk songs.) The similarity between vectors is multiplied by prior probabilities for each meter (based on the same data set). Whichever meter has the highest resulting "likelihood" (using the term fairly un-rigorously) is returned. Further processing, based on a beat similarity matrix, is then done to align the notes to beats, given the estimated meter.

Key Estimation

MusicIR can estimated both NoteQueue's key, and the chords implied by each note. To do this, it uses a hidden markov model (HMM). The observable outputs of the model are pitches. The hidden states of the model are the pitches' underlying/implied chords (24 of them. 12 pitch classes. Major and minor versions of each). Given some statistics about the probability of a chord occurring in a given key, and about the probability of one chord transitioning to another (both these sets of stats come from: http://research.microsoft.com/en-us/um/people/dan/chords/), and given some stats about the probability of a melody note occurring in the context of a certain chord in a certain key (made up. but good enough.), the HMM picks the key and implied chords that maximize the likelihood of observing the given set of melody notes. The HMM implementation is: https://github.com/dtkirsch/hmm.

Beat Detection

MusicIR does beat detection by calculating a song's Beat Similarity Matrix (see http://scholar.google.com/scholar?q=beat+similarity+matrix). BSMs are similar to autocorrelation. Where autocorrelation works on a list of scalars, similarity matrices work on a list of complex features (i.e. vectors). First, you define a similarity function for each vector element (pitch, duration, inteval, etc). You can use this to define a function that compares the similarity of any two beats. Finally, using that function, you can generate a 2D matrix of the similarity of each beat to every other beat. The strength of a given periodicity at various beat period (2 beats, 3 beats, etc) can be computed as the sum of the diagonal that many beats from the main diagnoal.

In addition to being a fairly effective means of detecting a fragment's time signature, beat similarity matrices are interesting as visualizations of rhythmic structure in songs: http://jimlindstrom.github.com/InteractiveMidiImproviser/