User’s Guide, Chapter 11: Corpus Searching

One of music21’s important features is its capability to help users examine large bodies of musical works, or corpora.

Music21 comes with a substantial corpus called the core corpus. When you download music21 you can immediately start working with the files in the corpus directory, including the complete chorales of Bach, many Haydn and Beethoven string quartets, three books of madrigals by Monteverdi, thousands of folk songs from the Essen and various ABC databases, and many more.

To load a file from the corpus, simply call corpus.parse and assign that file to a variable:

from music21 import *
bach = corpus.parse('bach/bwv66.6')

The music21 core corpus comes with many thousands of works. All of them (or at least all the collections) are listed on the Corpus Reference.

Users can also build their own corpora to index and quickly search their own collections on disk including multiple local corpora, for different projects, that can be accessed individually.

This user’s guide will cover more about the corpus’s basic features. This chapter focuses on music21’s tools for extracting useful metadata - titles, locations, composers names, the key signatures used in each piece, total durations, ambitus (range) and so forth.

This metadata is collected in metadata bundles for each corpus. The corpus module has tools to search these bundles and persist them on disk for later research.

Types of corpora

Music21 works with three categories of corpora, made explicit via the corpus.Corpus abstract class.

The first category is the core corpus, a large collection of musical works packaged with most music21 installations, including many works from the common practice era, and inumerable folk songs, in a variety of formats:

coreCorpus = corpus.corpora.CoreCorpus()


If you’ve installed a “no corpus” version of music21, you can still access the core corpus with a little work. Download the core corpus from music21’s website, and install it on your system somewhere. Then, teach music21 where you installed it like this:

>>> coreCorpus = corpus.corpora.CoreCorpus()
>>> coreCorpus.manualCoreCorpusPath = 'path/to/core/corpus'

Local Corpus

Music21 also can have one or more local corpora–bodies of works provided and configured by individual music21 users for their own research. They will be covered in Chapter 53. Anyone wanting to use them can jump ahead immediately to that chapter, but for now we’ll continue with searching in the core corpus.

localCorpus = corpus.corpora.LocalCorpus()

You can add and remove paths from a local corpus with the addPath() and removePath() methods:


Currently, after adding paths to a corpus, you’ll need to rebuild the cache.


We hope that this won’t be necessary in the future.

To remove a path, use the removePath() method.

 /Users/cuthbert/git/music21base/music21/corpus/ WARNING: local metadata cache: starting processing of paths: 0
 /Users/cuthbert/git/music21base/music21/corpus/ WARNING: cache: filename: /var/folders/mk/qf43gd_s5f30rzzbt7l7l01h0000gn/T/music21/local.p.gz WARNING: MetadataBundle Modification Time: 1629624939.985425 WARNING: Skipped 0 sources already in cache.
 /Users/cuthbert/git/music21base/music21/corpus/ WARNING: cache: writing time: 0.058 md items: 0

 /Users/cuthbert/git/music21base/music21/corpus/ WARNING: cache: filename: /var/folders/mk/qf43gd_s5f30rzzbt7l7l01h0000gn/T/music21/local.p.gz

By default, a call to corpus.parse or will look for files in any corpus, core or local.

Simple searches of the corpus

When you search the corpus, music21 examines each metadata object in the metadata bundle for the whole corpus and attempts to match your search string against the contents of the various search fields saved in that metadata object.

You can use to search the metadata associated with all known corpora, core, virtual and even each local corpus:

sixEight ='6/8')
 <music21.metadata.bundles.MetadataBundle {2163 entries}>

To work with all those pieces, you can parse treat the MetadataBundle like a list and call .parse() on any element:

myPiece = sixEight[0].parse()
 "I'll Touzle your Kurchy."

This will return a object which you can work with like any other stream. Or if you just want to see it, there’s a convenience .show() method you can call directly on a MetadataEntry.

You can also search against a single Corpus instance, like this one which ignores anything in your local corpus:

 <music21.metadata.bundles.MetadataBundle {2163 entries}>

Because the result of every metadata search is also a metadata bundle, you can search your search results to do more complex searches. Remember that bachBundle is a collection of all works where the composer is Bach. Here we will limit to those pieces in 3/4 time:

bachBundle ='bach', 'composer')
 <music21.metadata.bundles.MetadataBundle {363 entries}>'3/4')
 <music21.metadata.bundles.MetadataBundle {40 entries}>

Metadata search fields

When you search metadata bundles, you can search either through every search field in every metadata instance, or through a single, specific search field. As we mentioned above, searching for “bach” as a composer renders different results from searching for the word “bach” in general:'bach', 'composer')
 <music21.metadata.bundles.MetadataBundle {363 entries}>'bach', 'title')
 <music21.metadata.bundles.MetadataBundle {20 entries}>'bach')
 <music21.metadata.bundles.MetadataBundle {564 entries}>

So what fields can we actually search through? You can find out like this (in v2, replace corpus.manager with corpus.corpora.Corpus):

for field in corpus.manager.listSearchFields():

This field will grow in the future now that the development team is seeing how useful this searching method can be! Now that we know what all the search fields are, we can search through some of the more obscure corners of the core corpus:'taiwan', 'locale')
 <music21.metadata.bundles.MetadataBundle {27 entries}>

What if you are not searching for an exact match? If you’re searching for short pieces, you probably don’t want to find pieces with exactly 1 note then union that set with pieces with exactly 2 notes, etc. Or for pieces from the 19th century, you won’t want to search for 1801, 1802, etc. What you can do is set up a “predicate callable” which is a function (either a full python def statement or a short lambda function) to filter the results. Each piece will be checked against your predicate and only those that return true. Here we’ll search for pieces with between 400 and 500 notes, only in the core corpus:

predicate = lambda x: 400 < x < 500
corpus.corpora.CoreCorpus().search(predicate, 'noteCount')
 <music21.metadata.bundles.MetadataBundle {212 entries}>

You can also pass in compiled regular expressions into the search. In this case we will use a regular expression likely to find Handel and Haydn and perhaps not much else:

import re
haydnOrHandel = re.compile(r'ha.d.*', re.IGNORECASE)
 <music21.metadata.bundles.MetadataBundle {186 entries}>

Unfortunately this really wasn’t a good search, since we also got folk songs with the title of “Shandy”. Best to use a ‘*^*’ search to match at the beginning of the word only:

haydnOrHandel = re.compile(r'^ha.d.*', re.IGNORECASE)
 <music21.metadata.bundles.MetadataBundle {15 entries}>

We’ve now gone fairly high level in our searching. We will return to the lowest level in Chapter 12: The Music21Object