derekxkwan/syntheory_plus

Notes

• Dataset Viewer is disabled as we wanted to keep everything in WAV format with CSV metadata instead of Parquet files and the dataset is too large for Dataset Viewer to index properly.

Dataset Authors

Derek Kwan and Patrick Donnelly

Related Paper

The paper that introduces this dataset is "Probing for Advanced Music Theory Concepts in Generative Music Models" by Derek Kwan and Patrick Donnelly presented at EvoMUSART 2026

Dataset Creation

Rationale

This dataset was created to isolate specific Western music theoretical concepts in order to probe Meta's MusicGen and OpenAI's Jukebox music generative models (classification of extracted latent representations). The experimentation and more importantly here, the dataset, were based on the work of Wei et al. in their creation of the SynTheory dataset.

Dataset Creation

This dataset was creation following the process outline in Wei et al.'s "Do Music Generation Models Encode Music Theory?"". From programmatically generated MIDI files, we used the rustysynth library with the TimGM6mb soundfont.

Dataset Overview

This dataset consists of five subdatasets covering music theoretical concepts related to rhythm (polyrhythms), loudness (dynamics), and harmony (seventh chords, mode mixture, secondary dominants).

Each subdataset has its own dedicated folder with metadata.csv and consists of WAV file examples divided in class-balanced fold folders. There are 20 folds (folders 1-indexed) for each concept where the folds partition each concept to have an even distribution of classes across all folds (they were achieved using a stratified split). The metadata.csv consists of a manifest of all subdataset examples as well as associated metadata described in Dataset Fields.

In our paper introducing the dataset, we used the first 14 folds for training, the next three for validation, and the final three for testing in our classification experiments. This is reflected in the CSV field set_type.

We use 5 MIDI percussion instruments (Agogo, Woodblock, Taiko Drum, Melodic Tom, Snare Drum) to realize polyrhythms and dynamics subdatasets and 92 MIDI pitched instruments (same instruments as Wei et al.) to realize the seventh chords, mode mixture, and secondary dominants subdatasets. Each example is realized with one MIDI instrument (aside from polyrhythms, which is realized over pairs of MIDI instruments).

Following Wei et al., the polyrhythms and dynamics subdatasets are realized with three levels of reverb (0, 63, 127) as well as realized with different durations of offsets.

Each example is a 4 second, 44.1 kHz monophonic WAV file.

The harmony-related subdatasets are realized as quarter-notes in 4/4 at 60 bpm in closed voicing with a sustain of 90%. The seventh chords dataset realizes chords in a register closest to middle C (C4) and the mode mixture and secondary dominants datasets realize their chords where the root is closest to middle C.

Dataset Descriptions

Polyrhythms

This class-balanced dataset consists of 30,600 examples which are evenly distributed amongst 34 classes where classes are polyrhythm ratio realized by instrument pairs. That is, given two rates of speed r1 and r2 for two voices where r1 and r2 are coprime, a unique polyrhythm is defined by the ratio r1:r2. Here, we limit r1 and r2 to be between 2 and 11 and count r1:r2 and r2:r1 as the same polyrhythm.

Each polyrhythm is realized over one 4/4 bar at bpm 60 (1 bar in total), 120 (2 bars), and 180 (3 bars).

The offsets (delays in starting time) used here are 0, 120, 204, 700, and 1067 milliseconds.

Dynamics

This class-imbalanced dataset consists of 41,580 examples which are unevenly distributed amongst 7 classes where classes are dynamic expression pattern. Dynamic expression patterns here are hairpin (crescendo followed by decrescendo), reverse hairpin (decrescendo followed by crescendo, 8,505 examples), crescendo (2,835 examples), decrescendo (2,835 examples), subito forte (a subito louder dynamic, 8,505 examples), subito piano (a subito softer dynamic, 8,505 examples), and flat (no dynamic change, 1,890 examples). Aside from flat which is realized over one dynamic, dynamic patterns are realized over soft-loud pairs of dynamics where soft dynamics are [ppp,pp,p] and loud dynamics are [mf,f,ff]. Dynamics are derived from dividing the maximal MIDI value evenly by 6.

Each example is realized over a constant stream of notes in 4/4 at 60 bpm. This constant stream of notes occurs in note durations in subdivisions of the quarter note two through eight.

Hairpin, reverse hairpin, subito forte, and subito piano classes are characterized by inflection points where the dynamic or dynamic direction changes and thus are realized over inflection points at beats 2, 3, and 4.

The offsets (delays in starting time) used here are 0, 264, and 431 milliseconds.

Seventh Chords

This class-balanced dataset consists of 35,328 examples which are evenly distributed amongst 8 classes where classes are chord quality. The eight chord qualities of seventh chords contained in this subdataset are major 7th, minor 7th, major-minor 7th (also known as dominant 7th), minor-major 7th, half-diminished 7th, fully-diminished 7th, augmented with a major 7th (raised 5th, major 7th), augmented with a minor 7th (raised 5th, minor 7th).

Each seventh chord is realized in root, 1st, 2nd, and 3rd inversions. We acknowledge that in the case of the fully-diminished seventh that this may result in identical-sounding examples, but leave them in to retain class balance.

Mode Mixture

This class-balanced dataset consists of 19,872 examples which are evenly distributed amongst 2 classes where classes are the presence of mode mixture. The examples here are 4-chord triad major chord progressions taken from SynTheory and modified so that each IV chord is replaced by iv and each vi chord is replaced with bVI (built upon the lowered sixth scale degree).

The chord progression we adapt from SynTheory are I-IV-V-I, I-IV-vi-V, I-V-vi-IV, I-vi-IV-V, ii-V-I-vi, V-IV-I-V, V-vi-IV-I, vi-IV-I-V, and IV-I-V-vi.

The corresponding mode mixture chord progressions are I-iv-V-I, I-iv-bVI-V, I-V-bVI-iv, I-bVi-iv-V, ii-V-I-bVi, V-iv-I-V, V-bVI-iv-I, bVI-iv-I-V, and iv-I-V-bVI.

Secondary Dominants

This class-balanced dataset consists of 26,496 examples which are evenly distributed amongst 3 classes where classes are the presence of a secondary dominant or a tritone substitution of a secondary dominant (or lacking both, a diatonic chord progression). The examples here are 4-chord seventh chord major and minor chord progressions that follow circle-of-fifth root motion. We take diatonic chord progression and replace a chord with the secondary dominant of the following chord (for example, ii-V becomes V/V-V). We also replace the same chord with the tritone substitute of the secondary dominant of the following chord (ii-V becomes bII/V-V).

With M7 = major 7th, m7 = minor 7th, halfdim7 = half-diminished 7th, 7 = dominant 7th:

The major chord progressions are I(M7)-vi(m7)-ii(m7)-V7, I(M7)-I(M7)-IV(M7)-V7, I(M7)-ii(m7)-V7-I(M7), and I(M7)-iii(m7)-vi(m7)-V7.

The secondary dominant major chord progressions are I(M7)-V7/ii-ii(m7)-V7, I(M7)-V7/IV-IV(M7)-V7, I(M7)-V7/V-V7-I(M7), and I(M7)-V7/vi-vi(m7)-V7.

The tritone substitute major chord progressions are I(M7)-bII7/ii-ii(m7)-V7, I(M7)-bII7/IV-IV(M7)-V7, I(M7)-bII7/V-V7-I(M7), and I(M7)-bII7/vi-vi(m7)-V7.

The minor chord progressions are i(m7)-VI(M7)-ii(halfdim7)-V7, i(m7)-i(m7)-iv(m7)-V7, i(m7)-ii(halfdim7)-V7-i(m7), and i(m7)-III(M7)-VI(M7)-V7.

The secondary dominant minor chord progressions are i(m7)-V7/ii-ii(halfdim7)-V7, i(m7)-V7/iv-iv(m7)-V7, i(m7)-V7/V-V7-i(m7), and i(m7)-V7/VI-VI(M7)-V7.

The tritone substitute minor chord progressions are i(m7)-bII7/ii-ii(halfdim7)-V7, i(m7)-bII7/iv-iv(m7)-V7, i(m7)-bII7/V-V7-i(m7), and i(m7)-bII7/VI-VI(M7)-V7.

Dataset Fields

Polyrhythms

• name (str): The filename of the musical example (without the .wav extension)
• inst1 (int): The name of the first MIDI instrument used to realize the musical example
• inst2 (int): The name of the second MIDI instrument used to realize the musical example
• bpm (int): The BPM (beats per measure) the musical example is realized at
• num_bars (int): The number of 4/4 bars contained in the musical example. Each iteration of a polyrhythm occurs over the duration of one bar
• poly (str): The polyrhythm of the musical example in the format {first_rate}a{second_rate} (a for "against"). For example, 4a5 denotes "4 against 5"
• pair (str): The pair of MIDI instruments used in the musical example in the format {inst1}_{inst2}
• ratio (float): The ratio of the rates in the polyrhythm achieved by first_rate/second_rate. For example, 4a5 has a ratio of 0.8
• norm_ratio (float): The ratio of the rates mapping the smallest ratio of the dataset to 0 and the largest to 1. This is achieved with the expression ((ratio) - (smallest_ratio))/((largest_ratio)-(smallest_ratio))
• offset_lvl (int): Denotes an offset in starting time of the musical example in the WAV file. There are five possible 0-indexed offsets with values in milliseconds 0, 120, 204, 700, and 1067. These values were generated using a Numpy random number generator.
• offset_ms (int): The milliseconds in offset as described by offset_lvl.
• rvb_lvl (int): One of three levels of reverb applied to the MIDI realization. These are 0-indexed and correspond to MIDI values of 0, 63, and 127.
• rvb_val (int): The MIDI value of reverb as described in rvb_lvl
• poly1: The relative rate of the first MIDI instrument in the musical example
• poly2: The relative rate of the second MIDI instrument in the musical example
• polydist: The absolute value integer distance between the rates of the two instruments. For example, the polyrhythm of "5 against 11" has distance 6.
• label_idx: Each polyrhythm is mapped to a index (0-indexed) for downstream classification tasks.
• fold (int): The fold the musical example belongs in (1-indexed). The musical example can be found in the sub-dataset's fold_{fold} subfolder and there are 20 folds.
• set_type (str): Denotes if the musical example's fold is meant for training (train), testing (test), or validation (valid). For the original paper experiments, the first 14 folds are assigned to train, the next three to valid, and the next three to test.
• file_name (str): The relative path from the subdataset root folder of the WAV file (with .wav extension)

Dynamics

• name (str): The filename of the musical example (without the .wav extension)
• dyn1: One of two dynamics characterizing a dynamic expressive pattern. In non-flat patterns, this is the softer of two dynamics. In flat patterns, this is the same as dyn2.
• dyn2: One of two dynamics characterizing a dynamic expressive pattern. In non-flat patterns, this is the louder of two dynamics. In flat patterns, this is the same as dyn1.
• dyn_pair: A string formatted as {dyn1}-{dyn2}
• inst (str): The name of the MIDI instrument used to realize the musical example
• inflection_point: The beat of the 4/4 (1-indexed) bar where the dynamic direction changes for subp (drops in dynamic), subf (increases in dynamic), hairpin (changes from crescendo to decrescendo), and revhairpin (changes from decrescendo to crescendo)
• dyn_category: The type of dynamic expressive pattern (flat, decresc, cresc, subp, subf, hairpin, revhairpin)
• dyn_subcategory: The type of dynamic expressive pattern with inflection point appended with a - where applicable. For example, subf with an inflection point on beat 3 becomes subf-3.
• rvb_lvl (int): One of three levels of reverb applied to the MIDI realization. These are 0-indexed and correspond to MIDI values of 0, 63, and 127.
• rvb_val (int): The MIDI value of reverb as described in rvb_lvl
• offset_lvl (int): Denotes an offset in starting time of the musical example in the WAV file. There are three possible 0-indexed offsets with values in milliseconds 0, 264, and 431. These values were generated using a Numpy random number generator.
• offset_ms (int): The milliseconds in offset as described by offset_lvl.
• beats_per_bar (int): Number of beats per bar in a musical example. Each bar is in 4/4 time so this value of 4 for all musical examples.
• num_bars (int): Number of bars in each musical example. This is 1 for all musical examples.
• beat_subdiv (int): Denotes the number of equally-sized subdivisions that make up the stream of notes realizing each musical example. This value ranges from 2 (eighth-notes) to 8 (32nd notes).
• bpm (int): The BPM that a musical example is realized at. This is 60 BPM for all musical examples in this sub-dataset.
• num_beats (int): Number of beats in a musical example. As each musical example is at 60 BPM, this is 4 beats for all examples.
• fold (int): The fold the musical example belongs in (1-indexed). The musical example can be found in the sub-dataset's fold_{fold} subfolder and there are 20 folds.
• set_type (str): Denotes if the musical example's fold is meant for training (train), testing (test), or validation (valid). For the original paper experiments, the first 14 folds are assigned to train, the next three to valid, and the next three to test.
• file_name (str): The relative path from the subdataset root folder of the WAV file (with .wav extension)

Seventh Chords

• name (str): The filename of the musical example (without the .wav extension)
• inst (str): The name of the MIDI instrument used to realize the musical example
• root (str): The root of the seventh chord alongside its octave, with "s" denoting sharp. For example, "c4" is middle C while "as3" is the A# below that.
• pitch (str): The root of the seventh chord without its octave (just the pitch).
• octave (int): The octave of the root.
• quality (str): The quality of the seventh chord
• quality_idx (int): The quality of the seventh chord mapped to an index (starting with 0) for classification tasks.
• inv (int): The inversion of the seventh chord (0-indexed where root inversion is 0).
• bpm (int): The BPM a musical example is realized at. For all examples, this is 60 BPM.
• fold (int): The fold the musical example belongs in (1-indexed). The musical example can be found in the sub-dataset's fold_{fold} subfolder and there are 20 folds.
• set_type (str): Denotes if the musical example's fold is meant for training (train), testing (test), or validation (valid). For the original paper experiments, the first 14 folds are assigned to train, the next three to valid, and the next three to test.
• file_name (str): The relative path from the subdataset root folder of the WAV file (with .wav extension)

Mode Mixture

• name (str): The filename of the musical example (without the .wav extension)
• inst (str): The name of the MIDI instrument used to realize the musical example
• key_center (str): The pitch key center of a chord progrssion alongside its octave, with "s" denoting sharp. For example, "c4" is middle C while "as3" is the A# below that.
• scale_type (str): The primary mode that each example is realized in. As we only focus on the major mode, each musical example has the maj scale type.
• is_modemix (bool): false for no mode mixture in chord progression, true for mode mixture
• orig_prog (str): The roman numerals of each chord in integer scale degrees prefixed by the original chord progression's mode without mode mixture. For example, both major and mode mixture variants of "I-vi-ii-V" would have maj-1625 in this field.
• sub_prog (str): The same as orig_prog except mode mixture variants are now prefixed with mm1. Thus, the major variants of "I-vi-ii-V" would have maj-1625 in this field while the mode mixture variant would have mm1-1625.
• inv (int): The inversion (0-indexed) that each chord is in. As each chord is in root position for all musical examples, this is 0 for all.
• bpm (int): The BPM that each musical example is realized at. For all musical examples, this is 60 BPM.
• fold (int): The fold the musical example belongs in (1-indexed). The musical example can be found in the sub-dataset's fold_{fold} subfolder and there are 20 folds.
• set_type (str): Denotes if the musical example's fold is meant for training (train), testing (test), or validation (valid). For the original paper experiments, the first 14 folds are assigned to train, the next three to valid, and the next three to test.
• file_name (str): The relative path from the subdataset root folder of the WAV file (with .wav extension)

Secondary Dominants

• name (str): The filename of the musical example (without the .wav extension)
• inst (str): The name of the MIDI instrument used to realize the musical example
• key_center (str): The pitch key center of a chord progrssion alongside its octave, with "s" denoting sharp. For example, "c4" is middle C while "as3" is the A# below that.
• scale_type (str): The primary mode that each example is realized in. Major chord progressions are maj while minor chord progressions are min.
• sub_type (str): The type of chord progression that characterizes each musical example. N denotes a diatonic chord progression, S denotes a chord progression with a chord replaced by the secondary dominant of the following chord, and T deontes a similar replacement but with the tritone substitution of a secondary dominant.
• scale_sub_type (str): A string with the scale_type concatenated with the sub_type as {scale_type}_{sub_type}
• base_prog (str): The roman numerals of each chord in integer scale degrees.
• orig_prog (str): The base_prog string prefixed by the scale_type. This results in a string denoting the chord progression and the mode it occurs in. For example, "i7-VI(maj7)-ii(halfdim7)-V7" in minor is min-1625.
• inv (int): The inversion (0-indexed) that each chord is in. As each chord is in root position for all musical examples, this is 0 for all.
• sub_prog (str): The orig_prog string suffixed by the scale_sub_type. For example, a tritone substitution for "i7-VI(maj7)-ii(halfdim7)-V7" in minor is min-1625-T
• sub_pos (int): The position in the chord progression that a chord substitution occurs (1-indexed). For example, each chord progression that has a substitution in this sub-dataset substitutes the 2nd out of 4 chords and thus have the value 2. A chord progression where no substitution occurs (N types) have the value -1.
• bpm (int): The BPM a musical example is realized at. For all examples, this is 60 BPM.
• fold (int): The fold the musical example belongs in (1-indexed). The musical example can be found in the sub-dataset's fold_{fold} subfolder and there are 20 folds.
• set_type (str): Denotes if the musical example's fold is meant for training (train), testing (test), or validation (valid). For the original paper experiments, the first 14 folds are assigned to train, the next three to valid, and the next three to test.
• file_name (str): The relative path from the subdataset root folder of the WAV file (with .wav extension)