The kusc classical music dataset for audio key finding

The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.4, August 2014
THE KUSC CLASSICAL MUSIC DATASET FOR
AUDIO KEY FINDING
Ching-Hua Chuan1 and Elaine Chew2
1School of Computing, University of North Florida, Jacksonville, Florida, USA
2School of Engineering and Computer Science, Queen Mary, University of London,
London, UK
ABSTRACT
In this paper, we present a benchmark dataset based on the KUSC classical music collection and provide
baseline key-finding comparison results. Audio key finding is a basic music information retrieval task; it
forms an essential component of systems for music segmentation, similarity assessment, and mood
detection. Due to copyright restrictions and a labor-intensive annotation process, audio key finding
algorithms have only been evaluated using small proprietary datasets to date. To create a common base for
systematic comparisons, we have constructed a dataset comprising of more than 3,000 excerpts of classical
music. The excerpts are made publicly accessible via commonly used acoustic features such as pitch-based
spectrograms and chromagrams. We introduce a hybrid annotation scheme that combines the use of title
keys with expert validation and correction of only the challenging cases. The expert musicians also provide
ratings of key recognition difficulty. Other meta-data include instrumentation. As demonstration of use of
the dataset, and to provide initial benchmark comparisons for evaluating new algorithms, we conduct a
series of experiments reporting key determination accuracy of four state-of-the-art algorithms. We further
show the importance of considering factors such as estimated tuning frequency, key strength or confidence
value, and key recognition difficulty in key finding. In the future, we plan to expand the dataset to include
meta-data for other music information retrieval tasks.
KEYWORDS
Audio Key Finding, Ground Truth, Tuning, Dataset, Evaluation
1. INTRODUCTION
In this paper, we present a publicly available dataset for audio key finding. In tonal music, the key
establishes the context in which the roles of pitches and chords are defined. Determining key
from audio recordings is one of the most important music information retrieval (MIR) tasks; key
finding forms an essential component of systems for music segmentation, similarity assessment,
and mood detection. Due to issues such as copyright restrictions and the lack of annotated ground
truth, algorithms are often evaluated using relatively small datasets from private collections. In
addition, tonal complexity varies significantly in different genres and stylistic periods. Therefore,
simply using correct rates and dataset size to judge the performance of an audio key finding
system is a far from ideal approach. A commonly accessible dataset consisting of adequate
examples with sufficient variety and accurate annotations is thus needed.
The KUSC dataset presented in this paper consists of 15-second excerpts of compositions by
Bach, Mozart, and Schubert, including symphonies, concertos, preludes, fugues, sonatas, quartets
and more. We focus on the global key by examining only the first and last 15 seconds of the
DOI : 10.5121/ijma.2014.6401 1

recording—typically when the key of the piece is established and when the piece returns to the
original key, respectively—to reduce the possibility of key modulations. A common key labelling
method simply uses the key in the title, such as Symphony in G major, as ground truth.
While highly efficient, we show that this method can sometimes be prone to error. We introduce a
hybrid key annotation scheme that combines the use of title keys with selective manual validation
and correction of challenging cases. The manual annotations by expert musicians are
accompanied by ratings of key recognition difficulty, on a 5-point Likert scale, that provide
valuable information for evaluation of key-finding algorithms. Meta-data related to
instrumentation are also listed with the excerpt as tags.
For copyright reasons, the excerpts are shared in the form of commonly used and irreversible
acoustic features instead of audio waves. The acoustic features include the constant-Q
spectrogram, chromagram, and other mid-level features proposed by existing algorithms.
With the goal to develop the dataset as a benchmark for audio key finding, we conducted a series
of experiments using existing algorithms to provide an initial baseline for performance
comparisons. We use the dataset to test four existing audio key finding algorithms and report their
individual and combined scores for key finding accuracy. We further demonstrate that it is
important to consider in the evaluations factors such as tuning frequency, key
strength/confidence, and key recognition difficulty.
The recorded ensemble’s or instrument’s tuning directly impacts an algorithm’s chance of
success. Instrumentalists and ensembles may tune to reference frequencies different from A440;
for example, early music ensembles may tune to an A of 415 Hz, which would currently be heard
as an Ab. We first apply two tuning estimation methods and compare the difference between their
tuning frequency outputs. We also study the difference in estimated tuning between the first and
last 15 seconds of the same piece.
Using straightforward methods for determining key strength or confidence from the respective
models, we assessed the relation between the algorithm’s accuracy and the confidence or key
strength value. Based on the combined score, we investigate the relationship between key
recognition difficulty and instruments/music styles, and the algorithms’ accuracy.
Although the dataset is originally designed for audio key finding, we plan to expand the dataset
with more data for other MIR tasks such as automatic chord recognition and instrument
identification.
The remainder of the paper is organized as follows: Section 2 provides a general introduction to
audio key finding systems and current methods for evaluating them; Section 3 describes the
KUSC Classical music dataset, and the meta-data provided with the dataset; Section 4 introduces
the hybrid approach to key annotation and the rating of key-finding difficulty; Section 5 presents
the benchmark evaluations of four state-of-the-art audio key finding algorithms using the KUSC
dataset, followed by the conclusions.
2
2. AUDIO KEY FINDING AND RELATED WORK
2.1. Problem Definition and Background
The task of audio key finding is to identify the key of a music composition from its audio
recording. Polyphonic music comprise of multiple streams of notes sounded simultaneously, such
as music with multiple parts played simultaneously by different instruments as shown in the score

in Figure 1. Each part is a sequence of note and silent events, and each note event can be
described by a vector of properties, including <pitch, onset time, and duration>. For example, the
first three note events in the part that shown on the top of the score in Figure 1 are <
, 1, 0.5,
) consists of pitch class (bb) and register (4) that indicates the height of the pitch.
3
, 1.5, 0.125 and

, 1.625, 0.125, in which onset time and duration are presented in beats.
A pitch (
There exists 12 pitch classes including {c, c#/db, d, d#/eb, e, f, f#/gb, g, g#/ab, a, a#/bb, b}.
In Western tonal music, the key refers to the pitch context of a piece of music, and is represented
by the tonal center (a.k.a. the tonic), the most stable pitch class (for example, A or B) in the key.
The tonal center provides a reference for the identity and function of all other pitches in the piece.
For example, in a piece of Eb major key as shown in Figure 1, pitch eb is the tonic and pitch bb is
the fifth (five scale steps up from the pitch eb), the two most stable pitches in the piece. In
contrast, pitches such as a and b are less likely to appear in a piece of Eb major. Therefore, the
occurrence of the 12 pitch classes is an important indicator for the key. There are 24 keys in
Western tonal music, 12 major keys and 12 minor keys. In this study we concentrate on the global
key, the one key that operates over the entire length of a piece of music.
2.2. A General Architecture of Audio Key Finding Systems
A general architecture of audio key finding systems is illustrated in Figure 1. The input of the
system is an audio recording of a music composition or improvisation, and the output is the key.
An audio key finding system generally consists of two components: the first component focuses
on audio signal processing to extract features related to the pitch classes content of the sample,
while the second component determines the key based on the extracted pitch class information.
Figure 1. A general architecture of audio key finding systems
In the audio signal-processing component, the musical signal is first pre-processed by removing
silence and reducing noise. It is then divided into overlapped frames for spectral analysis. The
purpose of spectral analysis on short-duration frames is to obtain local information about note
events, particularly pitch information. Pitch information is related to the frequency of the signal.
In the equal temperament system, the interval between any two pitches corresponds to a simple
ratio with whole numbers indicating the relation between their frequencies. Using the standard
tuning 440Hz for pitch A4, the frequency of the pitch p that is n semitones away from pitch A4
can be calculated by:
Frequency
= 440 × 2

,
where n is a positive integer if the pitch is higher than A4 and a negative integer if it is lower.
The result of spectral analysis is further converted into a distribution representing the occurrence
of the 12 pitch classes. Various approaches have been proposed for the generation of pitch class

The International Journal of Multimedia Its Applications (IJMA) Vol.6, No.4, August 2014
distributions with the goal of increasing the accuracy of pitch information. A detailed survey of
existing algorithms is summarized in [1].
The key finding component in Figure 1 consists of representation models for the 24 keys and an
algorithm for determining the key by comparing the generated pitch class distribution and the
representation models. The representation model is usually defined as a typical pitch class
distribution for a specific key. The model can be constructed based on psychoacoustic
experiments [2], music theory [3], or learned from a dataset [4]. The comparison between the
generated pitch class distribution and the representation model in a key finding algorithm usually
involves calculating a distance function [5], correlations [6], or posterior probabilities [4]. The
key that has the shortest distance, or the highest correlation or probability, is chosen as the output.
2.3. Related Work
Table 1 summarizes existing approaches for audio key finding with their reported accuracy and
datasets used for evaluations.
4
Table 1. A summary of existing approaches for audio key finding
Approach Accuracy Dataset, Ground Truth, and Availability
Chai Verco [7] 84% 10 classical piano pieces
Ground truth was manually annotated by the author.
Chuan Chew [5, 8] 75% (410 classical),
70% (Chopin)
410 classical pieces / 24 Chopin preludes
Title key was used as ground truth.
Gómez Herrera [6] 64% 878 excerpts of classical music
zmirli [9] (template-based)
86% 85 classical pieces
zmirli [10] (local
key)
84%(pop), 76.9%
(classical)
17 pop / 17 classical excerpts
Ground truth was manually annotated.
Peeters [4] 81% 302 classical pieces
Papadopoulos
Peeters [1]
80.21% (classical),
61.31% (pop)
2 datasets:
• 5 movements of Mozart piano sonatas; ground truth
(key and chord progression) was manually
annotated; info. about composer and composition
title is available
• 16 real audio songs annotated by IRCAM
(QUAERO project)
Schuller Gollan
[11]
78.5% (pop),
88.8 (classical),
63.4 (jazz)
520 pieces in pop, jazz, classical and others
Ground truth was labeled by three musicians without
disagreements.
Tracks with key modulation were removed.
Sun, Li Ma [12] 68.7% in average 228 pieces in classical, pop, jazz, new age, and folk
Ground truth was labeled either by students or from
published music books.
Weiss [13] 92.2%, 93.7%, and
97%
3 datasets:
• 115 tracks of 29 symphonies: title key was used as
the ground truth; info. about composer and
symphony no. is available
• Saarland Music Data [14]: 126 tracks performed by
students; ground truth was manually annotated;
recordings are available in mp3
• 237 piano sonatas [15]: title key was used as the
ground truth; info. about composer, composition
title, and performer is available.

3. THE KUSC CLASSICAL MUSIC DATASET
This section describes the KUSC classical music dataset for audio key finding, focusing on
acoustic features, meta-data, and data formats that are made available with the dataset.
3.1. Music in the Dataset
The dataset used in this study was provided by Classical KUSC, a classical public radio station.
We selected compositions by Bach, Mozart and Schubert for audio key finding because they
represent three different styles with distinguishable levels of tonal complexity. In addition,
tonality is more clearly defined in these pieces than in more contemporary compositions. The
selected compositions include symphonies, concertos, sonatas, quartets, preludes, fugues, and
other forms with instruments including the recorder, violin, flute, piano, oboe, guitar, and choir.
For multi-movement works, we used only the first and last movement because these segments are
generally in the title key.
The entire dataset consists of 3224 different excerpts. We extracted two excerpts from each
recording: one containing the first 15 seconds and the other representing the last 15 seconds of the
recording. We only considered the 15-second segments of the recording because the key is more
likely to remain in the global key without modulations in these parts of the piece. Silence in the
beginning and the end of each recording were identified and removed using root-mean square
energy as described in [16].
The audio excerpts and their titles are not available for public access for copyright reasons.
However, low-level acoustic features related to audio key finding are made available for key
finding research as described in the following section.
3.2. Acoustic Features
Acoustic features related to tuning frequency and to presence of pitch classes are the most
important for audio key finding algorithms. Therefore, for each excerpt, we extracted frame-by-frame
features—the constant-Q spectrogram, chromagram, and pitch class profiles—related to
pitch classes using the fuzzy analysis center of effect generator algorithm proposed by Chuan and
Chew [5], Gómez’s harmonic pitch class profile [17], Lartillot and Toiviainen’s Matlab
implementation [16] of Gómez’s method, and Noland and Sandler’s hidden Markov models for
tonal estimation [18]. The features are given based on the tuning frequency estimated by two
existing methods [19, 20], as well as the standard tuning frequency (440 Hz for pitch A4).
5
3.2.1. Estimated Tuning Frequency
Here, we describe the two methods employed for estimating the tuning frequency.
The first is based on a method described by Müller and Ewert in [19]. The global tuning of a
recording is estimated by shifting the center of multi-rate filter banks to obtain an average
spectrogram based on several tuning frequencies. The deviation of the center that produces the
highest average value in the spectrogram is then identified as the estimated tuning frequency. In
this study, six deviation classes corresponding to shifts of {-1/2,-1/3,-1/4,0,1/4,1/3} semitones
from 440Hz were used.
The second is based on the technique outlined by Mauch and Dixon in [20]. Estimation of the
tuning frequency corresponds to the phase angle of discrete Fourier transform using circular
statistics. The frequency is wrapped onto the interval [-, ) where represents a quartertone.
Details of obtaining the estimated frequency can be found in [21].

Note that equal temperament is assumed in both tuning methods. To compare the differences
between the estimated tuning frequencies from the two methods, the output of [20] is mapped to
one of the categories in [19]: {427.4741, 431.6092, 433.6918, 440, 446.3999, 448.5539} Hz for
A4.
3.2.2. Pitch Class Distribution/Chromagram and Pitch-Based Spectrogram
Each pitch-based spectrogram consists of a series of vectors that represent spectral information
related to pitches as defined by their frequency ranges. For example, a pitch-based spectrogram
can be a matrix in which each row consists of 88 values for pitches A0 to C8. Pitch class
distributions or chromagrams, usually computed by folding values in pitch-based spectrograms
from pitches into pitch classes, are vectors of 12 values that relate to the distribution of 12 pitch
classes. Four algorithms [5, 16, 17, 18], to be outlined in Section 4.1.1, were used to extract
features for pitch-based spectrograms and pitch class distribution/chromagrams. In [5, 16, 17],
Fast Fourier Transform or Short Time Fourier Transform is used to produce spectral information.
In [18], the constant-Q transform is used. We briefly outline the concepts underlying these
methods.
In [17], Gómez detected spectral peaks using 120 bins per octave, a higher number of bins than
the standard 12 per octave for pitches. A triangular weighting function is used to reduce boundary
errors in the adjacent frequency bins. The resulting Harmonic Pitch Class Profile (HPCP) is a
vector of 120 values for 12 pitch classes. The chromagram in Lartillot and Toiviainen’s MIR
Toolbox [16] is another implementation of Gómez’s approach with slightly different default
parameter settings including frame size, hop size, pitch range and sampling frequency.
Chuan and Chew’s fuzzy analysis technique in [5] also generates a 12-value Pitch Class Profile
(PCP) for audio key finding. The technique uses knowledge of the harmonic series to reduce the
errors in noisy low frequency pitches. Pitches in low frequency are important to key finding
because they usually indicate the root of the chord played by the bass. However, it is difficult to
identify these pitches correctly because of the logarithmic scale in pitch frequency.
We also used Queen Mary University of London’s Vamp plug-in [22] to generate a constant-Q
spectrogram [23] for each excerpt. The spectrogram is used in Noland and Sandler’s key finding
method in [18]. The constant-Q spectrogram is the pitch-based spectrogram used to create the
folded chromagram.
Figure 2 provides an example of the extracted acoustic features for a recording of Bach’s Prelude
BWV 552 as shown in (a). The color specifies the significance of the observed frequency or
pitch, with the color of red for high values and blue or black for lower values. The resolution in
the grids depends on the algorithm’s default setting in parameters such as frame size and hop size.
Figure 2 (b) and (d) are generated by QM Vamp plug-in and (e) and (f) are generated using
Matlab, which result in similar color scheme for the figures using the same software. The y-axis
represents the 12 pitch classes in (d), (e), and (f) while the y-axis in (b) lists a range of detected
pitches. In Figure 2 (c), the y-axis is the 120 bins for the 12 pitch classes, which result in the
different visualization from other figures in Figure 2.
6

7
Figure 2. (a) The first three bars of Bach Prelude in Eb major BWV552, and extracted acoustic
features: (b) spectrogram [22], (c) HPCP [17], (d) chromagram [23], (e) mirchroma [16] and (f)
fuzzy analysis [5].
3.3. Meta-data
In addition to acoustic features and manual annotations, we also provided meta-data related to the
instrumentation. The meta-data consist of the tags, listed in Table 2, associated with each piece of
music.
Two types of information were manually annotated for the excerpts: ground truth key and key
recognition difficulty. We took a hybrid approach to ask musicians to manually annotate the key
only for the challenging excerpts without exhausting the annotators. We also noted the key
recognition difficulty by asking another three musicians to indicate how difficult it was for them
to determine the key using a 5-point Likert scale. The details for annotating ground truth key and
key recognition difficulty are described in Section 4.
For researchers to systematically compare the performance of their newly developed algorithm
with existing ones, we computed an average weighted score for each excerpt using the four audio
key finding algorithms [5, 16, 17, 18]. For each excerpt, each algorithm receives a score based on
the relation between their output key and the ground truth: 1 point (a full score) if the output is
identical to the ground truth, 0.5 if they are a perfect fifth apart, 0.3 if one is the relative
major/minor of the other, and 0.2 if one is the parallel major/minor of the other. An average is
then calculated using the four scores obtained by the algorithms. The details of the score
calculation and experiments are described in Section 5.
Table 2. Tags and their associated number
1 Soprano 7 Choir 13 Symphony 19 Harpsichord 25 Clarinet
2 Alto 8 Cello 14 Orchestra 20 Fortepiano 26 Oboe
3 Tenor 9 Violin 15 Concerto 21 Piano 27 Trumpet
4 Bass 10 Viola 16 Lute 22 Sonata 28 Bassoon
5 Baritone 11 String quartet 17 Guitar 23 Recorder 29 Horn
6 Chorus 12 Organ 18 Harp 24 Flute 30 Posthorn

3.4. Data Formats
The dataset is available via www.unf.edu/~c.chuan/KUSC/. Acoustic features described in Section
3.2.2 are pre-computed using different settings for the pre-defined parameters: frame size, hop
size, pitch range, sampling frequency, number of frequency bins per octave. The dataset can be
downloaded as delimited text files and MySQL database files.
4. ANNOTATING GROUND TRUTH AND KEY RECOGNITION DIFFICULTY
This section introduces our procedure to obtaining accurate ground truth key information for the
dataset by first identifying challenging musical fragments, then manually annotating them with
key information and key finding difficulty ratings.
4.1. A Hybrid Approach to Ground Truth Annotations
We first assigned the title key as ground truth for every excerpt in the dataset. To avoid
exhaustive manual examination of the entire dataset, we implemented five audio key finding
systems and used them to focus the manual examination on a subset consisting of challenging
excerpts in which the title key may not be the key. Each excerpt was first supplied as the input to
the five systems, which produces five key outputs. If no more than two out of the five systems
reported the same answer as the title key, the excerpt was labeled as a challenging case requiring
further manual examination. This challenging set was re-examined by three professional
musicians and their keys manually labelled by them.
4.1.1. Key Finding Algorithms
In this section, we briefly describe the audio key finding systems implemented for the
construction of the challenging set in detail, emphasizing the uniqueness of each system. More
details can be found in [24]. The methods we have chosen are mainly template-based approaches,
in which a pitch class distribution or chromagram is compared to a template key representation.
Note that systems that rely on training data were not implemented because the title key (ground
truth) may not be the key of the excerpt.
The five systems implemented include algorithms using Krumhansl and Schmuckler’s (K-S)
probe tone profile [2], Temperley’s modified version of the K-S model [3], zmirli’s template-based
correlation model [9], Gómez’s harmonic pitch class profile (HPCP) method [17], and our
fuzzy analysis center of effect generator (FACEG) algorithm [5]. The first four algorithms all
compute a correlation coefficient between the pitch class distribution of the excerpt and the
template pitch class profile for each key. They differ mostly in their template pitch class profiles
as shown in Figure 3. We thus implemented a basic approach to generate pitch class distributions
using the Fast Fourier Transform.
In [5], we proposed an audio key finding system called fuzzy analysis center of effect generator.
A fuzzy analysis technique is used for PCP generation using the harmonic series to reduce the
errors in noisy low frequency pitches. The PCP is further re-fined periodically using the current
key information. The representation model used in the system is Chew’s Spiral Array model [25,
26], a representation of pitches, chords, and keys in the same 3-dimensional space with distances
reflecting their musical relations. The Center of Effect Generator (CEG) key finding algorithm
determines key via a nearest-neighbor search in the Spiral Array space in real-time: an
instantaneous key answer is generated in each window based on past information. The CEG is the
only model that considers pitch spelling.
8

9
Figure 3. C major and C minor key profiles proposed by Krumhansl, Temperley, zmirli, and Gómez.
Figure 4. The flowchart of the hybrid approach for ground truth annotations
4.1.2. Experiments and Results
The process of examining the accuracy of the title key and re-labeling the ground truth is
illustrated in Figure 4. An excerpt was first examined by the five implemented audio key finding
systems. If more than two systems out of the five reported the key identical to the one in the title,

we saved the title key as the ground truth. If not, the excerpt was moved to the challenging set and
re-examined by two musicians. If the two musicians had the same key for the excerpt, then we
saved the re-labeled key as the ground truth. If the two musicians disagreed, the excerpt was
examined by the third musician. Experiment results were reported by comparing the accuracy of
the five audio key finding systems using the title key and the relabelled key as ground truth.
Out of the 3324 excerpts, 727 excerpts (21.87%) were moved to the challenging set that was
examined by musicians. Table 3 lists the distribution of the challenging set, in absolute numbers
and as a percentage of the number of excerpts we considered by each composer.
10
Table 3. Details of the entire dataset and the challenging set by Bach, Mozart, and Schubert
Composer Total # of recordings Challenging set (first 15 sec) Challenging set (last 15 sec)
Bach 553 245 (44.30%) 244 (44.12%)
Mozart 873 75 (8.59%) 98 (11.23%)
Schubert 236 24 (10.17%) 41 (17.37%)
Figures 5 and 6 show the experiment results using the title key and the re-labeled key
respectively. The reported key are divided into nine categories based on its relation to the ground
truth: correct (Cor), dominant (Dom), sub-dominant (Sub), parallel major/minor (Par), relative
major/minor (Rel), same mode with the root one semitone higher (M+1), same mode with the root
one semitone lower (M-1), same mode but not in the previous categories (MO), and relations not
included in any of the previous categories (Others).
Figure 5. Key finding results for the challenging dataset using the title key as ground truth.

11
Figure 6. Key finding results for the challenging dataset using the re-labeled key as ground truth.
Comparing Figure 5 (a), (b) with Figure 6 (a), (b), it can be observed that most of the incorrect
answers using the title key fall into the last three categories, especially (M-1), which indicates that
tuning may be an issue in key finding for Bach’s pieces. Similar results can be observed in
Mozart’s pieces by comparing Figure 5 (c), (d) with Figure 6 (c), (d). However, the significant
proportion in parallel major/minor (Par) category in Figure 5 (d) indicates that many Mozart
pieces actually end in the parallel key with respect to the title key. In Figure 5 (e), the reported
keys are more evenly distributed among the categories for Schubert’s pieces using the title key as
the ground truth. This implies more complex compositional strategies that the composer
employed to start the piece. The result in Figure 5 (f) for the last 15-second of Schubert’s pieces
is more similar to Mozart’s than Bach’s, showing that the pieces end in the parallel major/minor
key or a key that has the same mode as the title key. More detailed analysis can be found in [24].
4.2. Key Recognition Difficulty Annotations
In addition to reporting the key of the excerpts, we also collected data about key recognition
difficulty. Three musicians were asked to listen to the excerpts in the challenging dataset, and use
the 5-point Likert scale, with 1 indicating the easiest level and 5 the most difficult, to rate the
difficulty that they experienced in recognizing the key. One objective of this data collection step
was to observe how individuals perceive this difficulty and the difference between individuals’
responses. Therefore, no formal definition of key recognition difficulty, nor examples of each
level, were given to the musicians.
We also studied the consistency of individual rater’s annotations by examining the rater’s
responses to multiple versions of the same compositions, i.e., recordings of the same piece by
different performers. A consistent rater should indicate similar difficulty levels and identical, or
closely related, keys for the different recordings of the same piece.

Figure 7 shows the rater’s consistency according to two measures: (a) the weighted average of
standard deviations of rated difficulty levels; and, (b) the ratio of number of distinct keys to total
number of duplicate excerpts. The first measure is the average of the standard deviations,
computed for each set of multiple recordings of the same piece, weighted by the number of
recordings in the set. This measure is bounded between zero and 2.5, when two scores are at the
extreme ends. The second measure serves as a confusion indicator for determining the key for
different versions of the same piece. For the given dataset, this measure is bounded between
0.248 and 1, when every duplicated excerpt is assigned a different key. According to both
measures, the raters’ labels are found to be consistent as the results are closer to the lower bounds
of each measure. We also proposed ways to automatically predict key recognition difficulty based
on acoustic features as described in [27].
12
Figure 7. Consistency of individual rater’s labels
5. EVALUATIONS USING THE KUSC DATASET
After we collected data for the ground truth and key recognition difficulty, we used the dataset
and the annotated data to test several state-of-the-art audio key-finding algorithms as described in
this section. The test focuses on five aspects: tuning, audio key finding accuracy against re-labeled
ground truth, the relation between an algorithm’s accuracy and its confidence value, the
relation between an algorithm’s accuracy and human perceived key recognition difficulty, and the
relation between an algorithm’s accuracy and instruments/musical styles. The experiment’s
results not only provide a detail examination of the dataset, but also offer a baseline to which any
newly developed algorithm can be compared in the future.
5.1. Tuning
Figure 8 shows the difference, measured in number of semitones, between the estimated tuning
frequencies generated by the Chroma Toolbox [19] and non-negative least squares (NNLS) [20].
It can be observed that the outputs of the two tuning estimation methods are mostly the same
(2961 out of 3224 excerpts). However, the two methods reported different estimated tuning
frequencies, in some cases more than 0.8 semitones apart.
In addition to examining the difference between the outputs of the two tuning methods, we also
compared the estimated tuning frequencies of the first and last 15 seconds of the same
composition. Table 4 lists the percentage of excerpts in which the estimated tuning frequencies
are different for the beginning and ending segments. The percentage in Table 4 is relatively high
based given that the tuning usually stays the same throughout a recording. More studies will be
conducted to examine the reasons for the observed discrepancies.

13
Figure 8. The distance between the tuning frequencies estimated using the Chroma Toolbox [19] and
NNLS [20]
Table 4. The percentage of the estimated tuning frequencies that differ in the first and last 15 seconds
Tuning method Bach Mozart Schubert
Chroma toolbox [19] 12.45% 23.22% 17.45%
NNLS [20] 10.65% 21.49% 17.02%
5.2. Audio Key Finding Accuracy
We tested four audio key finding algorithms [5, 16, 17, 18] using the dataset. Since it is not the
focus of this paper to optimize the algorithms, we only implemented the basic versions of these
algorithms. We used their default settings for parameters such as frame size, hop size, and pitch
range to generate frame-by-frame pitch class profiles or chromagrams. We then calculated the
average profile over all frames.
For the algorithms described in [16, 17], the average profile was then compared to the pre-defined
templates for the 24 keys, and the template with the highest correlation value was then reported as
the key. For FACEG [5], the only algorithm that considered pitch spelling, we re-moved the
requirement of pitch spelling by considering all spellings for the purpose of comparison. The
average profile was used to generate the Center of Effect (CE) representing the tonal center in the
3-dimensional Spiral Array model [1]. The key was determined by searching for the key
representation, from among the pre-defined points representing the 24 keys, closest to the CE.
The algorithm in [18] constructs a hidden Markov model with 24 major and minor key states and
observations representing chord transitions. The key profiles in the implementation of the
algorithm are created from analysis of Bach’s Well Tempered Clavier, Book I. The implemented
program available via QM Vamp plugin [22] tracks local keys and key changes. For comparison
with the other approaches, we needed to select one key from the multiple local keys as the global
key for the excerpt. The selection was based on the total duration of the key, i.e., the local key
reported with the longest period was chosen to be the global key.
To prevent tuning frequencies that deviated from the norm from affecting the key finding result,
we used the output of the tuning estimation methods to specify the reference frequency for the
algorithms instead of fixing it at 440Hz for A4. For conciseness, in this paper, we only report the
results using the estimated frequency from the Chroma Toolbox [19].
Table 5 shows the key finding accuracy, overall weighted scores, for the four algorithms. Each
weighted score is calculated based on the relation of the output key and the ground truth key (as
described in Section 3.4), and converted to a percentage by dividing by the total number of
excerpts. Note that the key finding program in the MIR Toolbox [16] reported the highest
weighted score, with the HPCP [17] reporting one very close to it. The low percentage in the
result of [18] may be due to the fact that the program is too sensitive to local changes. The

availability of the KUSC dataset will allow researchers to further investigate the effect of local
key modulations on global key determination.
For each excerpt, we calculated the average of the four scores from the four algorithms. Figure 9
shows the distribution of average weighted scores for the dataset. A peak can be observed at 0.7,
with low values on either side. This information is helpful for evaluating a new audio key finding
algorithm: excerpts that receive an average score above 0.7 can be used as benchmarks for
evaluating the basic performance of a key finding algorithm.
14
Table 5. The four algorithms’ key finding accuracy scores for the KUSC classical music dataset
MIR HPCP FACEG QM
Weighted score %* 91.7% 90.34% 82.64% 59.74%
*identical = 1, dominant = 0.5, relative = 0.3, parallel = 0.2
Figure 9. The histogram showing the distribution of average weighted scores of the four algorithms
5.3. Key Strength and Confidence Value
We also used the dataset to examine the relation between key strength/confidence value and key
determination accuracy. In addition to the output key, audio key finding algorithms also report a
numeric value corresponding to key strength or confidence, a value indicating the certainty of the
output key. Such key strength/confidence values have also been used by zmirli as the weights in a
voting scheme to determine the global key from multiple local keys [9].
For the key finding algorithm in HPCP [17] and the MIR Toolbox [16], we used the correlation
between the pitch class profile and the 24 keys as the key strength value. The confidence value in
FACEG [5] is calculated using the distance between the CE and the two closest keys. Suppose d1
is the distance between the CE and the closest key and d2 is the same for the second closest key.
The confidence value is calculated as (d2 – d1)/ d1, the scaled difference between the distance to
the closest and second closest key. The key detector program in the QM Vamp plugin [22]
outputs key probabilities for all 24 keys for each frame. In this study, the key strength value is
given by the average of the global key’s probabilities across all frames.
Table 6 lists the correlation between key strength/confidence value and accuracy as represented
by the weighted score. Although some algorithms use this value to determine key, the value is not
strongly correlated with the accuracy of the output key.
The correlation coefficients between the key strength/confidence values of pairs of the four
algorithms are shown in Table 7. Except for the higher correlation between the key finding
programs in the MIR Toolbox and HPCP, both based on the Krumhansl-Schmuckler template-

matching algorithm [2], no strong correlations are observed between any other pairs of
algorithms.
15
Table 6. Correlation between the key strength/confidence value and accuracy scores for the four
algorithms.
MIR HPCP FACEG QM
Correlation 0.2666 0.2922 0.092 -0.0592
Table 7. The four algorithms’ key finding accuracy scores for the KUSC classical music dataset
MIR HPCP FACEG QM
MIR 1 - - -
HPCP 0.5479 1 - -
FACEG 0.1124 0.0653 1 -
QM 0.0633 -0.0144 -0.0030 1
5.4. Key Recognition Difficulty
To understand the connection between challenges faced by the algorithms and musicians, we
calculated the correlation between the algorithms’ average weighted accuracy score and the
annotators’ average key recognition difficulty ratings. Although the musicians were instructed to
use the 5-point scale for their ratings, they may not use the scale the same way. Therefore, in
addition to simply calculating the average of the three ratings, we also normalized and
standardized their ratings before calculating the correlation coefficient.
Table 8 shows the correlation results. The values in the table show negative correlation between
the difficulty ratings and the average weighted score generated by the four algorithms. However,
the correlation is not strong.
Table 8. Correlation between the key recognition difficulty and weighted accuracy scores.
Unprocessed
difficulty
Normalized
difficulty
Standardized
difficulty
Correlation -0.3026 -0.2817 -0.2970
Since the correlation in Table 8 was not strong, we further conducted t-tests. Based on the
distribution of the average weighted scores as shown in Figure 9, we propose a null hypothesis
(H0) that excerpts with an average weighted score less than 0.7 received the same mean difficulty
ratings as those with scores equal to or above 0.7. The alternative hypothesis (H1) is that excerpts
with a score less than 0.7 have a significant lower mean difficulty rating than ones with scores
equal to or above 0.7. Results of the t-tests, shown in Table 9, reject the null hypothesis. Thus
excerpts receiving a weighted accuracy score higher than 0.7 from the algorithms do have a mean
difficulty rating lower than those with weighted accuracy scores below 0.7.
Table 9. The result of the t-tests
Unprocessed
difficulty
Normalized
difficulty
Standardized
difficulty
H 1 1 1
p-value 1.4808 x 10-8 1.2505 x 10-7 2.0356 x 10-8

5.5. Instrumentation
For excerpts that contain instrument and music style tags as listed in Table 2, we calculated the
mean of average weighted scores for each tag to examine how instrumentation impacts the
algorithms’ accuracy. The result is shown in Figure 10 and it indicates that wind instruments (tags
no. 23 – 30), except for the oboe (tag no. 26), have higher average weighted scores than others. It
is also observed that excerpts labelled as vocal tracks (tag no. 1 – 7) report higher standard errors
than others. However, more data is needed in order to obtain more conclusive results.
16
Figure 10. The average weighted scores with standard errors for the 30 instrumentation tags listed in
Table 2.
6. CONCLUSIONS
A dataset consisting of 3224 classical music excerpts for audio key finding was described in the
paper. To share the dataset with the research community without violating copyright, excerpts
were represented as commonly used and irreversible acoustic features such as chromagrams and
constant-Q spectrograms instead of audio source files. In addition, manual annotations such as
ground truth for the global key and key recognition difficulty were collected. Meta-data relating
to instrumentation was also made available.
We conducted a series of experiments with the dataset using several algorithms pertaining to
aspects of the audio key finding process. First, for tuning frequency estimation, the two methods
tested agreed on the same frequency for over 90% of the excerpts. But 10% to 23% of the
beginning excerpts returned a different tuning frequency than that for the corresponding end
portion of the same piece. Four audio key finding algorithms were also tested with weighted
accuracy scores ranging from 60% to 92%. In addition to the estimated key, many algorithms also
produce a numerical value indicating the strength or confidence of the output key. We did not
observe clear relations between this key strength/confidence value and the algorithm’s accuracy.
We also analyzed the relations between the four algorithms’ key finding accuracy and musicians’
key recognition difficulty ratings. Although accuracy was not strongly correlated with difficulty
rating, the result showed that the mean key difficulty rating of excerpts receiving weighted scores
above 0.7 is significantly lower than that for excerpts with weighted accuracy scores lower than
0.7. Finally, we examined the relations between the algorithms’ accuracy and instruments tags.
Some patterns were observed, but these were not sufficient for drawing strong conclusions.
For future work, we plan to expand the dataset by including more excerpts and generating
features and annotations for other music information retrieval tasks.

REFERENCES
[1] Papadopoulos, H. Peeters, G. (2012) “Local key estimation from an audio signal relying on
harmonic and metrical structures,” IEEE Transaction on Audio, Speech, and Language Processing,
Vol. 20, No. 4, pp. 1297–1312.
[2] Krumhansl, C. L. (1990) “Quantifying tonal hierarchies and key distances,” Cognitive Foundations of
17
Musical Pitch, chapter 2, pp. 16–49, Oxford University Press, New York.
[3] Temperley, D. (1999) “What’s Key for Key? The Krumhansl-Schmuckler Key-Finding Algorithm
Reconsidered,” Music Perception, Vol. 17, No. 1, pp. 65–100.
[4] Peeters, G. (2006) “Musical key estimation of audio signal based on HMM modeling of chroma
vectors,” Proceedings of the International Conference on Digital Audio Effects, 127-131.
[5] Chuan, C.-H. and Chew, E. (2005) “Fuzzy analysis in pitch class determination for polyphonic audio
key finding,” Proceedings of the 6th International Conference on Music Information Retrieval.
[6] Gómez, E. and Herrera, P. (2004) “Estimating the tonality of polyphonic audio files: Cognitive versus
machine learning modeling strategies,” Proceedings of the 5th International Conference on Music
Information Retrieval, 92-95.
[7] Chai, W. and Verco, B. (2005) “Detecting of key change in classical piano music,” Proceedings of the
6th International Conference on Music Information Retrieval, 468-473.
[8] Chuan, C.-H. and Chew, E. (2006) “Audio key finding: considerations in system design, and case
studies on 24 Chopin’s preludes,” Ichiro Fujinaga, Masataka Goto, George Tzanetakis (eds),
EURASIP Journal on Applied Signal Processing, Special Issue on Music Information Retrieval.
[9] zmirli, Ö. (2005) “Template based key finding from audio,” Proceedings of the International
Computer Music Conference.
[10] zmirli, Ö. (2007) “Localized key finding from audio using non-negative matrix factorization for
segmentation,” Proceedings of the 8th International Conference on Music Information Retrieval, 195-
200.
[11] Schuller, B. Gollan, B. (2012) “Music theoretic and perception-based features for audio key
determination,” Journal of New Music Research, Vol. 41, No. 2, pp. 175–193.
[12] Sun, J., Li, H., Ma, L. (2011) “A music key detection method based on pitch class distribution
theory,” International Journal of Knowledge-based and Intelligent Engineering Systems, Vol. 15, pp.
165–175.
[13] Weiss, C. (2013) “Global key extraction from classical music audio recordings based on the final
chord,” Proceedings of the Sound and Music Computing Conference, pp. 742–747.
[14] Müller, M., Konz, V., Bogler, W. and Arifi-Müller, V. (2011) “Saarland music data,” Proceedings of
the 12th International Conference on Music Information Retrieval.
[15] Pauws, S. (2004) “Musical key extraction from audio,” Proceedings of the 5th International
Conference on Music Information Retrieval.
[16] Lartillot, O. Toiviainen, P. (2007) “Matlab toolbox for musical features extraction from audio,”
Proceedings of the 10th International Conference on Digital Audio Effects, pp. 237–244.
[17] Gómez, E. (2006) Tonal Description of Music Audio Signals, PhD thesis.
[18] Noland, K. Sandler, M. (2007) “Signal processing parameters for tonality estimation,” Proceedings
of Audio Engineering Society 122nd Convention.
[19] Müller, M. Ewert, S. (2011) “Chroma toolbox: matlab implementations for extracting variants of
chroma-based audio features,” Proceedings of the 12th International Society for Music Information
Retrieval Conference, pp. 215–220.
[20] Mauch, M. Dixon, S. (2010) “Approximate note transcription for the improved identification of
difficult chords,” Proceedings of the 11th International Society for Music Information Retrieval
Conference, pp. 135–140.
[21] Mauch, M. (2010) Automatic Chord Transcription from Audio Using Computational Models of
Musical Context, PhD thesis, Queen Mary University of London.
[22] www.vamp-plugins.org/plugin-doc/qm-vamp-plugins.html#qm-keydetector, accessed July 2014.
[23] Brown, J. (1991) “Calculation of a constant Q spectral transform,” Journal of the Acoustical Society
of America, Vol. 89, No. 1, pp. 425–434.
[24] Chuan, C.-H. and Chew, E. (2012) “Generating ground truth for audio key finding: When the title key
may not be the key,” Proceedings of the 13th International Conference on Music Information
Retrieval.
[25] Chew, E. (2000) “Towards a mathematical model of tonality,” doctoral dissertation, Department of
Operations Research, Massachusetts Institute of Technology, Cambridge, Mass, USA.

[26] Chew, E. (2001) “Modeling tonality: applications to music cognition,” Proceedings of the 23rd
18
Annual Meeting of the Cognitive Science Society, pp. 206–211.
[27] Chuan, C.-H. and Charapko, A. (2013) “Predicting key recognition difficulty in polyphonic audio,”
Proceedings of the 9th IEEE International Workshop on Multimedia Information Processing and
Retrieval.
AUTHORS
Ching-Hua Chuan is an assistant professor of computing at Universityof North
Florida. She received her Ph.D. in computer science from University of Southern
California Viterbi School of Engineering in 2008. She received her B.S. and M.S.
degrees in electrical engineering from National Taiwan University. Dr. Chuan’s
research interests include audio signal processing, music information retrieval,
artificial intelligence and machine learning. She was the recipient of the best new
investigator paper award at the Grace Hopper Celebration of Women in Computing
in 2010.
Elaine Chew received the B.A.S. degree in mathematical and computational
sciences with honors, and in music with distinction, from Stanford University,
Stanford, California, in 1992, and the S.M. and Ph.D. degrees in operations
research from the Massachusetts Institute of Technology, Cambridge,
Massachusetts, in 1998 and 2000, respectively, in the United States. She joined
Queen Mary, University of London, in the United Kingdom, as Professor of Digital
Media in Fall 2011, where she serves as Director of Music Initiatives in the Centre
for Digital Music. Her research interests center on mathematical and computational
modelling of music prosody and structure, on which she has authored over 80
refereed journal and conference articles. Prof. Chew was the recipient of a National Science Foundation
Faculty Early Career Award, Presidential Early Career Award in Science and Engineering, and Edward,
Frances, and Shirley B. Daniels Fellowship at the Radcliffe Institute for Advanced Study.
.

The kusc classical music dataset for audio key finding

More Related Content

Viewers also liked (18)

Similar to The kusc classical music dataset for audio key finding (20)

Recently uploaded (20)

The kusc classical music dataset for audio key finding