Feature definition

The following table presents the definition of all the stock features provided with musif.

The first column are Python Regular Expressions that match the relevant columns of the final DataFrame.

The .+ at the beginning of a pattern stands for the scope prefix of the column: Part<Abbreviation>_ (one part), Sound<Abbreviation>_ (all parts of one instrument), Family<Abbreviation>_ (all parts of one family), or Score_ (all parts). Each row states the scopes at which the feature is actually produced. Staves that resolve to the same part abbreviation are treated as one logical part (counts summed, extremes merged). For a sound performed by exactly one part, every Sound<X>_ column equals the corresponding Part<X>_ column by construction (the sound scope aggregates over that one part); the DataProcessor option delete_duplicated_sound_columns drops these duplicates, keeping the per-part means and part counts.

Conventions used throughout:

• every _Per column is a proportion between 0 and 1, not a percentage;

• a value that cannot be computed is NaN, never 0: a ratio whose denominator is zero, a statistic of nothing, any feature of a part that is not in the score. A 0 is always a measured value. The replace_nans option of the DataProcessor keeps to this: it turns an empty cell into 0 only in the columns that count events or give their proportion, and only for the scores in which the part has notes (and, for the columns that depend on the harmonic analysis, in which the analysis is available);

• pitches are analysed as they sound: the part of a transposing instrument (horn, trumpet, clarinet…) is transposed by the interval stated in its file before any feature is computed, so its scale degrees and ambitus, and the aggregates that include it, describe the notes that are heard. A part whose file states no transposition is taken as written;

• intervals and melodic motion follow melodic lines. In each part, the measures written in one voice and the first voice of the measures written in several voices form the main line. Every other voice is a line of its own, and it starts anew at each passage (a run of consecutive measures in which that voice is written): nothing is measured from one passage to the next. A measure in which nothing at all is written for the part (the second part of a split staff, while the staff shows a single line) ends the main line in the same way;

• chords (double stops, keyboard chords) are excluded from note-based features, although a measure containing only chords still counts as sounding;

• scale degrees in minor keys are measured against the natural minor scale (the signature that the key has by modern convention, which is not always the one notated), so a raised leading tone is Degree#7. A degree column therefore does not mean the same in the two modes: Degree7 is the leading tone of a major key and the subtonic of a minor one. The same holds, passage by passage, for the degrees relative to the local key. Compare arias across modes with Mode at hand;

• rates are given in two units of time. The beat is the one of each time signature (NumberOfBeats: four in 4/4, two in 2/2 and 6/8, one in 3/8), and a rate per beat (Density, RhythmInt, DottedRhythm, DynGrad, DynAbruptness, HarmonicRhythmBeats) changes with the metre a piece is notated in: the same notes written in 2/2 give twice the density per beat they give in 4/4. The columns ending in Quarters give the same rates per quarter note, which is the same length in every metre; use them to compare arias in different metres. Neither unit knows the tempo;

• by default every feature describes the notated score, read once through: repeats, da capo and dal segno instructions are not followed. In an aria whose reprise is not written out, the opening section is on the page once although it is performed twice, so proportions and means weigh it less than a score with a written-out reprise does. RepriseInstruction tells the two kinds of score apart. With expand_repeats, every feature describes instead the score as it is played: its measures are unfolded in the order that ms3 gives its MuseScore file (repeat signs, voltas, da capo and dal segno to Fine or Coda), and the harmonic annotations follow the same order. This needs the MuseScore file, and it is applied only where that file is the MusicXML file annotated, bar for bar (as many measures, each as long); a score that cannot be unfolded is analysed as written, and RepeatsUnfolded says which is which. Where the order jumps, the measure jumped to keeps the time signature, key signature and clef in force at its place;

• the Score_ value of a feature covers every part of the score, and depends on which parts those are. Where it pools the notes of all parts (intervals, scale degrees, ambitus), it is dominated by the parts with most notes, usually the strings. Where it is a mean over parts (Density, RhythmInt, DottedRhythm, the dynamics), a pair of horns that plays twenty notes weighs as much as the first violins, and an aria with winds comes out less dense than the same music without them. To compare arias of different scoring, use the columns of the parts they share (PartVnI_, PartBs_, FamilyStr_…) rather than Score_;

• in windowed extraction only complete windows are analyzed (a tail shorter than one window is skipped) and per-window rows have Measures equal to the window size.

Finally, the Stock module column lists the module distributed in musif that generates each feature. The stock feature modules are: core, ambitus, melody, tempo, density, texture, harmony, lyrics, scale, scale_relative, key, dynamics, rhythm and music21, plus the basic modules scoring and file_name_generic.

Note that none of these features works on percussion instruments.

In addition to the features in the present page, the features in the 'music21' stock module (first row of the below table) provide all the features from music21.features.native.

Column RegEx	Explanation	Stock module
m21_.*	Music21 features from music21.features.native. The QualityFeature extractor (P22) is excluded: it duplicated Mode and read only notated key signatures, contradicting the actual mode. The column set is identical with and without caching	music21
Id	Numeric identifier of the file within the extraction	added by the extractor
WindowId	Index of the window within its score (0 when no windowing is used)	added by the extractor
WindowRange	Half-open range of 0-based measure indices covered by the window, e.g. 0 - 8	added by the extractor
FileName	File name of the original score	core
Artist	Text before the first underscore of the file name (empty when there is none)	file_name_generic
Title	Text after the first underscore of the file name (the whole stem when there is none)	file_name_generic
Measures	No. of measures of the piece (of the window, in windowed extraction)	core
.+_SoundingMeasures$	No. of measures that have at least one note (not all silence). Part scope counts the part’s own measures; sound/family/score scopes are sums over their parts	core
.+_SoundingMeasuresMean$	Sounding measures per part, i.e. the summed sounding measures / no. of parts, at sound/family/score scope	core
.+_Notes$	No. of notes. Part scope counts the part; sound/family/score scopes are sums over their parts	core
.+_NotesMean$	Notes per part, i.e. the summed notes / no. of parts, at sound/family/score scope	core
TimeSignature	Initial time signature of the piece (later changes are not reported here)	tempo
TimeSignatureGrouped	Type of meter of the initial time signature: simple duple (1/X, 2/X, 4/X, 8/X), simple triple (3/2, 3/4, 3/8, 3/16), compound duple (6/X), compound triple (9/X), compound quadruple (12/X), or other. 3/8 is simple triple, although NumberOfBeats counts one beat per measure in it	tempo
NumberOfBeats	No. of beats per measure implied by the initial time signature	tempo
Key	Key in which the piece is written. The relation to annotated key signature may not be straightforward	key
KeySignature	Key signature of the piece according to modern convention; might not correspond to the one annotated on the score	key
FamilyInstrumentation	Comma-separated abbreviations of the instrumental (non-voice) families present on the score	scoring
FamilyScoring	Comma-separated abbreviations of all families present on the score, voices included	scoring
SoundScoring	Comma-separated sound abbreviations of everything present on the score, voices included	scoring
Instrumentation	Instruments on the score, using Italian abbreviation; 2 parts of the same instrument are simplified as 1 such instrument	scoring
Scoring	List of parts on the score, including voices, using Italian abbreviation; 2 parts of the same instrument are not simplified. A part carries a Roman numeral when its own name does (Violino II), or when it is one of the two players split from a shared staff; staves left with the same abbreviation (the two sopranos of a duet) are one part and appear once each	scoring
.+_NumberOfParts$	Number of parts found on the score, for each sound and family and for the whole score. Staves are counted: a staff split into two players counts two, and two staves with the same abbreviation (the two sopranos of a duet) count two, although they are one logical part	scoring
.+_NumberOfFilteredParts$	Number of parts kept by parts_filter (equal to NumberOfParts when no filter is set), for each sound and family and for the whole score	scoring
Voices	Voice parts included on the score, ordered from high to low when the voice type is known. Voices of the same type are not numbered: they are one logical part	scoring
KeySignatureType	Type of KeySignature (b flats, n naturals, or s sharps) according to modern convention	key
Mode	Mode of the Key of the piece	key
NumericTempo	Numeric tempo of the piece, as stated by the engravers and correctors in the MusicXML file: the first <sound tempo> of the first measure of the first part. Empty when that measure states none, and when it states 120, the default of notation software, which cannot be told from a tempo set on purpose	tempo
Tempo	Verbal annotation for the tempo, as defined by the composer: the first tempo word of the first measure (of the first full measure, when the score opens with a pickup that carries none). Later changes of tempo are not reported	tempo
RepriseInstruction	1 when the score carries a da capo or dal segno instruction, 0 when it does not. With such an instruction the reprise is performed but not written out, so every proportion and mean describes the notated page, on which the opening section appears once; without it, a reprise that exists is on the page twice. Compare arias across this column with care	tempo
RepeatsUnfolded	1 when the measures of the score were unfolded in the order in which they are played (expand_repeats, see the conventions above), 0 when it is analysed as written: always without expand_repeats, and with it where the score has no MuseScore file or its MuseScore file is barred differently or does not encode the reprise the score asks for	tempo
TempoGrouped1	Classification of Tempo according to the main term in the verbal annotation	tempo
TempoGrouped2	Classification of Tempo in Slow, Moderate, or Fast (NA when no tempo term is recognized)	tempo
.+_Density$	Notes per beat: summed notes of the scope’s parts / summed per-measure beats of those parts (metre changes are respected at every scope, and a single-part sound equals its part). The beat is the one of each time signature (see NumberOfBeats): one per measure in 3/8, two in 2/2 and 6/8, so values are comparable only within one metre. Produced at part/sound/family/score scope	density
.+_DensityQuarters$	Notes per quarter note. The same per quarter note instead of per beat: the quarter note is the same in every metre, so this version is comparable between arias in 4/4, 2/2, 3/8…	density
.+_SoundingDensity$	Notes per sounding beat: summed notes / summed per-measure beats of the sounding measures, a ratio of sums over the scope’s parts. NaN for a scope that never sounds. Produced at part/sound/family/score scope	density
.+_SoundingDensityQuarters$	Notes per quarter note of the sounding measures. The same per quarter note instead of per beat: the quarter note is the same in every metre, so this version is comparable between arias in 4/4, 2/2, 3/8…	density
Harmony_Analysis_Available	1 when the harmonic features were successfully extracted from the MuseScore analysis file, 0 otherwise (no analysis file, or the extraction failed)	harmony
Harmony_Chord_.+_(Count|Per|PerTime)	No. / proportion of annotations / proportion of annotated time of each chord on a degree of the key in force, labelled by its numeral plus its chord type (V7, viio, …; all major/minor seventh qualities collapse to 7, and the leading-tone chords of minor keys, #viio, #viio7 and #vii%7, are counted with viio, viio7 and vii%7). Augmented sixths appear under their own names (It, Ger, Fr). Applied chords are not here but under Harmony_Applied_. _Per counts annotations (a chord held for four measures and a passing chord count the same); _PerTime weighs each annotation by its length. The denominators are all the chords of the piece, so Chord_ and Applied_ together sum to 1	harmony
Harmony_Applied_.+_(Count|Per|PerTime)	The same for the chords applied to a degree other than the tonic of the key in force (V7/V, viio7/ii…), labelled by what they are (V7, viio7) whatever degree they lead to	harmony
Harmony_Chord_types_.+	Proportion of harmonic annotations of each chord type: major triad, minor triad, 7th, dim, aug, aug6, other. With the suffix _PerTime, proportion of the annotated time	harmony
Harmony_Chords_Grouping_1.+_(Count|Per|PerTime)	No. / proportion of annotations / proportion of annotated time of the chords of each harmonic function in the key in force: T, SD, D, MED, NAP, ST, LN… Upper and lower case follow the quality of the chord (sd is a minor chord with subdominant function, such as ii in major), not the mode of the key. Groups of their own: rm (vi in a major key, the relative minor) and rj (III in a minor key, the relative major); SEC for applied chords, whose function lies in another key; A6 for augmented sixths. Every chord has the function of the numeral it is annotated with: a cadential six-four is a D where it is annotated V(64), and an I64 is a T even when a dominant follows. Chords whose label cannot be classified are excluded from both grouping levels	harmony
Harmony_Chords_Grouping_2.+_(Count|Per|PerTime)	The same without the distinction of chord quality: T, SD, D, MED, rel (rm and rj together), SEC, A6, NAP, ST… Note: Grouping_1 T/t/MED/med hold a single numeral each, and Grouping_2 NAP/bSD/#SD only merge the case pair of the same Grouping_1 bucket	harmony
Harmony_HarmonicRhythm	Harmonic rhythm of the piece, calculated as no. of harmonic annotations / Measures	harmony
Harmony_HarmonicRhythmBeats	Harmonic rhythm of the piece by beats, calculated as no. of harmonic annotations / total no. of beats (beats are summed over the measures of the score, each after the time signature in force in it, whether or not it carries a label). It follows the metre: an aria in 3/8 has one beat per measure, one in 4/4 four	harmony
Harmony_HarmonicRhythmQuarters	Harmonic annotations per quarter note. The same per quarter note instead of per beat: the quarter note is the same in every metre, so this version is comparable between arias in 4/4, 2/2, 3/8…	harmony
Harmony_Key_.+_PercentageMeasures	Proportion of measures that are in a given local key, extracted from the LocalKey column of the harmonic analysis. Measures are counted by their position in the file (a second ending and a bar excluded from the numbering count as measures), each key lasts from its first label to the label of the next key, and the last key lasts to the end of the piece	harmony
Harmony_Key_Modulatory_.+	Proportion of key areas (sections) spent in a given local key, extracted from the LocalKey column of the harmonic analysis	harmony
Harmony_Key_(D|d|Dom|SD|sd|SubD|T|rel|Other)_PercentageMeasures and Harmony_Key_Modulatory_(D|d|Dom|SD|sd|SubD|T|rel|Other)	Riemannian groupings of the two rows above (D=V, d=v, Dom=D+d, SD=IV/II/VI, sd=iv/ii, SubD=SD+sd, T=I/i, rel=III/vi), added by the DataProcessor when grouped_analysis is on. D, d, T and their Modulatory twins are single-member sums, so they duplicate the V/v/I-or-i rows they aggregate	postprocessing
Harmony_Numerals_.+_(Count|Per|PerTime)	No. / proportion of annotations / proportion of annotated time of each numeral of the key in force, as included in the “Numeral” column of the harmonic analysis. Applied chords are counted together under Applied, and augmented sixths under Aug6 (the analysis spells them as a vii or a V applied to the dominant): neither is a degree of the key	harmony
Harmony_Additions_.+	Proportion of annotations whose “changes” field falls in each bucket: 4_6_64 (suspension/six-four figures), +9, Others_+ (other added notes), Others_no_+ (other alterations). The altered note with which the analysis spells out an augmented sixth is not counted. Only non-empty buckets are emitted; the denominator is the total number of annotation rows, including the rows that carry only a phrase or pedal marking (the other harmony families count chords only)	harmony
Part.+_Spe_avg_abs_step_.+	Melodic speed, in semitones per quarter note. 1. split the part into melodic segments (a new segment starts after every rest and at every chord, which is not part of a melodic line, and each melodic line, as defined in the conventions above, is its own sequence of segments); 2. resample each segment on a regular grid of S quarter notes (each note covers its span of grid points); 3. take the differences between consecutive samples inside each segment — never across a rest — and divide by S; 4. this feature is the mean of the absolute differences pooled over all segments. Steps S: 0.25, 1 and 2 (SPEED_STEPS in musif.extract.features.melody.constants). NaN when no segment has enough samples. The time unit is the notated quarter note: the same music written in halved note values gives twice the speed. Part scope only	melody
Part.+_Asc_avg_step_.+_win_.+	1. smooth each resampled segment with a centered rolling mean of 2·W+1 samples; 2. a “chunk” is a maximal run of strictly rising smoothed samples within one segment; this feature is the average length of the rising chunks, in quarter notes, pooling the chunks of all segments. Steps S: 0.25, 0.5 and 1; windows W: 2 and 4 (CHUNK_STEPS, CHUNK_WINS), so the smoothing spans 1.25 to 9 quarter notes. Only segments that hold at least one full window are measured. NaN when there is no rising chunk, or no segment long enough. Part scope only	melody
Part.+_Dsc_avg_step_.+_win_.+	Like Asc_avg but for descending chunks. Part scope only	melody
Part.+_Asc_prp_step_.+_win_.+	Proportion of rising steps over all smoothed steps of the measured segments (see Asc_avg). NaN when no segment holds a full window. Part scope only	melody
Part.+_Dsc_prp_step_.+_win_.+	Proportion of falling steps over all smoothed steps of the measured segments. NaN when no segment holds a full window. Part scope only	melody
.+_AverageDuration$	Average duration of the note values (quarter note = 1, eighth note = 0.5, etc.), excluding zero-duration grace notes. Produced at part/score scope	rhythm
.+_DottedRhythm$	No. of single-dotted figures (a dotted note shorter than one beat followed, on the same beat of the same measure, by a shorter note) / the beats of the sounding measures, summed measure by measure after the time signature in force. Both the beat that a figure must fit in and the denominator depend on the metre. Produced at part/score scope (score = mean over parts with sounding measures)	rhythm
.+_DottedRhythmQuarters$	The same figures per quarter note of the sounding measures	rhythm
.+_DoubleDottedRhythm$	No. of doubly dotted figures (same rule, two dots) / the beats of the sounding measures, summed measure by measure after the time signature in force. Produced at part/score scope	rhythm
.+_DoubleDottedRhythmQuarters$	The same figures per quarter note of the sounding measures	rhythm
.+_RhythmInt$	Sum of the duration of the notes (in quarter notes) / the beats of the sounding measures, summed measure by measure after the time signature in force. It is the length of the beat times the share of the time that is filled with notes, so it mostly tells the metre. Produced at part/score scope (score = mean over parts)	rhythm
.+_RhythmIntQuarters$	Sum of the duration of the notes / quarter notes of the sounding measures: the share of the sounding time that is filled with notes (above 1 where two lines of a part sound together). Comparable between metres	rhythm
.+Degree(b+|#x?|x)?\d_Count$	No. of notes corresponding to the given scale degree in relation to the main key of the piece (natural minor reference in minor keys). Produced at part/score scope; the score is the sum over all parts	scale
.+Degree(b+|#x?|x)?\d_Per$	Proportion of the row above over the notes counted for the scope (NaN when the scope has no notes)	scale
.+Degree(b+|#|x)?\d_Count_relative	No. of notes corresponding to the given scale degree in relation to the local key of the passage the note is in (requires MuseScore harmonic analysis). A note takes the key of the last label at or before it, found by the position of its measure in the file and its onset in the measure (printed bar numbers differ between the MusicXML and the MuseScore file at Finale’s X bars and in voltas); before the first label it takes the first key, after the last the last. Produced at part/score scope; the score is the sum over all parts	scale_relative
.+Degree(b+|#|x)?\d_Per_relative	Proportion of the row above over the notes counted for the scope	scale_relative
.+_Degree_(Asc|Desc|Dasc|Ddesc|Nat|Nonat)_(Count|Per)	Grouped degrees relative to the main key: raised (#), lowered (b), doubly raised (x), doubly lowered (bb), natural, and non-natural degrees. Added by the DataProcessor when grouped_analysis is on, for every scope prefix that carries degree columns	postprocessing
.+_Degree_(Asc|Desc|Dasc|Ddesc|Nat|Nonat)_(Count|Per)_relative	Same grouping for the local-key degrees	postprocessing
.+_DynAbruptness$	Sum of absolute values of ((dynamic marking 2 − dynamic marking 1) / no. of beats from marking 1 to marking 2), divided by the total no. of beats. Produced at part/score scope	dynamics
.+_DynAbruptnessQuarters$	The same with both times in quarter notes instead of beats, comparable between metres	dynamics
.+_DynGrad$	Mean rate of change between successive markings: sum of absolute values of ((dynamic marking 2 − dynamic marking 1) / no. of beats from marking 1 to marking 2, rests included), divided by the no. of such changes. NaN with fewer than two levels. Produced at part/score scope	dynamics
.+_DynGradQuarters$	The same per quarter note instead of per beat, comparable between metres	dynamics
.+_DynMean$	Mean of the dynamic values of the part’s markings, each level counted once however long it lasts. A repeated identical marking counts once, and fp and pf contribute both levels. Accents are not levels: sf, sfz and fz are skipped, and sfp sets the level p. Two text marks set a level, “dolce” (p) and “sotto voce assai” (pp); other text marks (“sotto voce”, “mezza voce”, “dolce assai”, “cresc.”…) and hairpins are ignored. The values are DYNAMIC_VALUES in musif.extract.features.dynamics.constants. A silence is not a level: a marking stays in force through rests until the next one. NaN for parts without markings. Produced at part/score scope; the score value is the mean over the parts that carry markings, so it depends on which instruments the score has	dynamics
.+_DynMean_weighted$	Mean dynamic value over the sounding beats: sum of (dynamic value × no. of sounding beats in which it is in force) / no. of sounding beats that have a marking in force. What sounds before the first marking has no stated level and is left out. Produced at part/score scope (score = mean over the parts that carry markings)	dynamics
.+_Ambitus$	Ambitus (highest minus lowest MIDI pitch), in number of semitones. Produced at part/sound/family/score scope	ambitus
.+_HighestNoteIndex$	MIDI pitch of the highest note. Produced at part/sound/family/score scope	ambitus
.+_LowestNoteIndex$	MIDI pitch of the lowest note. Produced at part/sound/family/score scope	ambitus
.+_HighestNote$	Highest note, as note name with octave. Produced at part/sound/family/score scope	ambitus
.+_LowestNote$	Lowest note, as note name with octave. Produced at part/sound/family/score scope	ambitus
.+_IntervallicKurtosis$	Kurtosis of the interval sizes (difference between the MIDI pitch of two notes). Melody features are produced at part/sound/score scope (no family)	melody
.+_IntervallicStd$	Sample std of the interval sizes (difference between the MIDI pitch of two notes)	melody
.+_IntervallicMean$	Mean of the signed interval sizes (difference between the MIDI pitch of two notes). The signed intervals of a melodic line cancel out: their sum is the last pitch minus the first. This is therefore the net displacement of the lines divided by the number of intervals, close to 0 in any part, and it says little about the shape of the melody; see TrimmedIntervallicMean and the absolute statistics for that	melody
.+_IntervallicSkewness$	Skewness of the interval sizes (difference between the MIDI pitch of two notes)	melody
.+_AbsoluteIntervallicKurtosis$	Kurtosis of the absolute interval sizes (in no. of semitones)	melody
.+_AbsoluteIntervallicMean$	Mean of the absolute interval sizes (in no. of semitones)	melody
.+_AbsoluteIntervallicSkewness$	Skewness of the absolute interval sizes (in no. of semitones)	melody
.+_AbsoluteIntervallicStd$	Sample std of the absolute interval sizes (in no. of semitones)	melody
.+_AbsoluteIntervallicTrimDiff$	.+_AbsoluteIntervallicMean − .+_TrimmedAbsoluteIntervallicMean	melody
.+_AbsoluteIntervallicTrimRatio$	(.+_AbsoluteIntervallicMean − .+_TrimmedAbsoluteIntervallicMean) / .+_AbsoluteIntervallicMean (NaN when the mean is 0)	melody
.+_TrimmedAbsoluteIntervallicMean$	Mean of the absolute interval sizes (in no. of semitones), trimming the most extreme 10% at each tail	melody
.+_TrimmedAbsoluteIntervallicStd$	Sample std of the absolute interval sizes (in no. of semitones), trimming the most extreme 10% at each tail	melody
.+_TrimmedIntervallicMean$	Mean of the interval sizes (in no. of semitones, positive and negative), trimming the most extreme 10% at each tail	melody
.+_TrimmedIntervallicStd$	Sample std of the interval sizes (in no. of semitones, positive and negative), trimming the most extreme 10% at each tail	melody
.+_AscendingIntervallicMean$	Mean of the sizes of the ascending intervals (in no. of semitones); NaN when there are none	melody
.+_AscendingIntervallicStd$	Sample std of the sizes of the ascending intervals (in no. of semitones)	melody
.+_DescendingIntervallicMean$	Mean of the sizes of the descending intervals (negative, in no. of semitones)	melody
.+_DescendingIntervallicStd$	Sample std of the sizes of the descending intervals (in no. of semitones)	melody
.+_AscendingIntervals_Count$	No. of ascending intervals	melody
.+_AscendingIntervals_Per$	Proportion of .+_AscendingIntervals_Count over all intervals	melody
.+_DescendingIntervals_Count$	No. of descending intervals	melody
.+_DescendingIntervals_Per$	Proportion of .+_DescendingIntervals_Count over all intervals	melody
.+_Interval[A-Za-z]+-?\d+_(Count|Per)	No. / proportion of each specific interval, named by music21’s directed name (descending intervals carry a - before the number, e.g. IntervalM-3)	melody
.+_Intervals.+Asc_(Count|Per)	No. / proportion of ascending intervals of a given type (Perfect, Major, Minor, Augmented, Diminished, double variants, WithinOctave, BeyondOctave)	melody
.+_Intervals.+Desc_(Count|Per)	No. / proportion of descending intervals of a given type	melody
.+_Intervals.+All_(Count|Per)	No. / proportion of intervals of a given type in either direction (unisons included). Note: WithinOctaveAll_Per + BeyondOctaveAll_Per = 1, and the Asc (resp. Desc) within/beyond pair sums to AscendingIntervals_Per (resp. descending)	melody
.+_StepwiseMotion.+_(Count|Per)	No. / proportion of [Asc, Desc, All] stepwise intervals (seconds, and chromatic unisons such as A1/d1)	melody
.+_Leaps.+_(Count|Per)	No. / proportion of [Asc, Desc, All] leaps (interval of a 3rd or larger)	melody
.+_RepeatedNotes_Count$	No. of consecutively repeated notes (the pitch does not change: 0 semitones). Equals IntervalP1_Count except for enharmonic 0-semitone steps such as d2, which music21 names as seconds	melody
.+_RepeatedNotes_Per$	Proportion of .+_RepeatedNotes_Count over all intervals	melody
.+AscendingSemitones_Sum$	Sum of all the ascending intervals, counted as semitones	melody
.+DescendingSemitones_Sum$	Sum of all the descending intervals, counted as semitones (negative)	melody
.+_LargestAbsoluteSemitonesAll$	Largest interval in number of semitones, disregarding the direction (the interval of largest magnitude in either direction). The Asc/Desc variants were removed: they duplicated LargestSemitonesAsc/Desc up to sign	melody
.+_LargestSemitones(Asc|Desc|All)$	Largest [asc, desc, all] interval in number of semitones, keeping the sign. When the largest rising and falling intervals are equal (a +12 and a −12), All is the rising one	melody
.+_LargestInterval(Asc|Desc|All)$	Largest [asc, desc, all] interval, as the notated music21 interval name (All: the rising one when the two directions tie, as above)	melody
.+_SmallestSemitones(Asc|Desc)$	Smallest [asc, desc] interval in number of semitones, keeping the sign. The All and AbsoluteSemitones variants were removed: the smallest-magnitude interval degenerates to the unison whenever any note is repeated	melody
.+_SmallestInterval(Asc|Desc)$	Smallest [asc, desc] interval, as the notated music21 interval name	melody
.+_MeanInterval$	Interval name of the rounded mean absolute interval (it disregards direction, so it is never a descending interval)	melody
Score_VoiceReg	Where the centre of the vocal part lies in relation to the tonic: the distance in semitones from the mean MIDI pitch of the part to the nearest tonic of the key of the piece (between −6 and 6, positive when the centre lies above that tonic), averaged over the vocal staves (two staves of the same voice type count twice)	lyrics
.+_Syllables$	No. of syllables in a given vocal part, as introduced by the music engraver: the lyrics that carry text (a note carrying several verses contributes one syllable per verse; the text-less marker that closes a melisma line is not a syllable). Produced at part/score scope	lyrics
.+_SyllabicRatio$	No. of notes / no. of syllables. The score scope is the ratio of the summed notes and syllables of all vocal parts (a note-weighted aggregate). NaN without lyrics	lyrics
.+_VoicePresence$	SoundingMeasures / measures of the part, for each singer in the score	lyrics
Presence_of_.+	Inclusion (1/0) of a given part on the score; empty for a score that failed to extract	added if separate_instrumentation_column is true in the DataProcessor configuration (default: false)
Part.+|Part.+_Texture	No. of notes of Part[1] / no. of notes of Part[2], using each part’s actual note count. NaN when Part[2] has no notes. Pairs follow the configured scoring order, so a pair always yields the same column across a corpus	texture

Notes for typical notated corpora

On the opera-seria corpus used to validate these features (and most notated corpora of the period), several column families are structurally degenerate; they are kept for generality but carry no information there:

• DoubleAugmented/DoubleDiminished interval columns are all zero. DoubleDottedRhythm is not: the corpus has 411 doubly dotted notes in 70 scores, but it is zero in most arias.

• m21_ChordBassMotionFeature_0 is the constant 1.0 sentinel that music21 emits when a score contains no chord-symbol objects (notated scores never do); features _1 through _11 are zero.

• the processed Degree_.+_Dasc/_Ddesc columns are zero, so Nonat = Asc + Desc there.

• rare chromatic degrees and interval qualities are all-zero on most corpora; wherever a directional count is all-zero, its All twin equals the other direction.

• Presence_of_X is constant 1 for instruments present corpus-wide (voices are excluded from Instrumentation by design), and equals “Sound<X>_NumberOfParts is not NaN” row by row.