Conference PaperPDF Available

Towards the Identification of Irish Traditional Flute Players from Commercial Recordings


Abstract and Figures

This paper explores whether distinct spectral differences exist between professional flute players of Irish traditional music and between their playing of different notes and whether we could identify players based on their spectral differences in commercial recordings. The audio signal is represented using short-term magnitudes of the first few harmonic frequencies. Player identification is performed by employing the Gaussian classifier, where each player is characterised using a single multivariate Gaussian distribution with full covariance matrix. Experimental evaluations are performed using audio recordings from five professional flute players. Only the sustained sections of notes, which were manually identified based on note onsets, are used. The identification of players is explored for each note separately and using various number of note instances.
Content may be subject to copyright.
Islah Ali-MacLachlan, M
unevver K
uer, Cham Athwal
School of Digital Media Technology,
Birmingham City University, UK
{islah.ali-maclachlan, munevver.kokuer,
Peter Jan
School of Electronic, Electrical
& Systems Engineering
University of Birmingham, UK
This paper explores whether distinct spectral dierences exist be-
tween professional flute players of Irish traditional music and
between their playing of dierent notes and whether we could
identify players based on their spectral dierences in commer-
cial recordings. The audio signal is represented using short-term
magnitudes of the first few harmonic frequencies. Player identifi-
cation is performed by employing the Gaussian classifier, where
each player is characterised using a single multivariate Gaussian
distribution with full covariance matrix. Experimental evalua-
tions are performed using audio recordings from five professional
flute players. Only the sustained sections of notes, which were
manually identified based on note onsets, are used. The identifi-
cation of players is explored for each note separately and using
various number of note instances.
Irish Traditional Music is a melodic form of instrumen-
tal music that was originally used to accompany dancing.
It has, along with many forms of folk music, undergone
a revival and is enduringly popular for players and listen-
ers alike. Informal ‘sessions’, whilst a relatively modern
phenomenon, allow flute players to gather with players of
other traditional and modern melody and rhythm instru-
ments including fiddle (violin), tin whistle, mandolin, ac-
cordion, guitar and bouzouki (Williams, 2010). Whilst
new music is constantly being added to the tradition, the
melodies collected by O’Neill (1998) and first published
in 1907 still form a central corpus.
The most popular style of flute used by traditional play-
ers is the wooden concert D flute. It evolved from the sim-
ple system unkeyed flute and is available in both unkeyed
(diatonic) and keyed (chromatic) styles. It continued to be
used by traditional players long after Boehm’s metal con-
cert flute became the norm for orchestral players (Hamil-
ton, 1990). One of the main reasons for this, as noted by
Breathnach (1996) is that ornaments such as sliding into
notes and rolling, a very fast inter-note articulation, are not
possible with Boehm’s key design.
Playing a flute requires control over breath pressure,
lip position and lip aperture (Coltman, 1968). Changes in
these parameters will result in timbral dierences, a con-
tributing factor to tonal dierences between players. Tim-
bre is the quality of a musical note that distinguishes it
from other notes of the same pitch and loudness. Schouten
(1968) proposed five acoustic parameters to describe tim-
bre: Range between tonal and noiselike character, spectral
envelope, changes in spectral envelope and fundamental
frequency over time, sound onset and time envelope.
Erickson (1975) agrees that these dimensions are excel-
lent for perceptual analysis of music. Handel (1995) argues
that there are many stable and time-varying acoustic prop-
erties but no single property fully defines timbre. Burred
et al. (2010) found that temporal envelope is more valu-
able when distinguishing between sustained and decaying
instruments. In this study we do not analyse notes based
on temporal attributes.
According to Jensen (1999) spectral envelope is often
enough to identify a sound. Timoney et al. (2004) found
that in spectral magnitude plots of tin whistles, a closely
related instrument to the flute, there are only a small num-
ber of harmonics compared to other instruments. Chudy &
Dixon (2010) presented a cello player recognition system
employing timbre features. They used controlled studio
recordings of 6 cellists playing the same excerpt on two
dierent cello instruments.
Utilising spectral analysis, our earlier study showed clear
timbral dierences between players with varying levels of
experience, each playing single G
notes as part of a scale
with dierent models of wooden flute (Ali-MacLachlan
et al., 2013). We found that magnitudes of F
, F
and F
varied the most between individual players. This work fol-
lowed on from Widholm et al. (2001) who identified tim-
bral dierences between individual classical flautists play-
ing Boehm system flutes manufactured from a range of
metals. Both studies showed that individual players have
significantly more eect on timbre than changes in mate-
rial or pattern.
To build on our earlier study (Ali-MacLachlan et al.,
2013), we explore whether distinct spectral dierences ex-
ist between professional players and between their playing
of dierent notes and whether we could identify profes-
sional players based on their spectral dierences in com-
mercial recordings. We analyse the playing of five players,
all at a professional level, playing reels, jigs, hornpipes and
polkas. These are some of the most common forms of Irish
traditional dance tunes. We investigate timbre-identity of
players across the notes that have the highest usage in the
corpus, specifically, D
, E
, F#
, G
, A
, B
, D
, E
. The analysis is performed using the sustained sections
of the notes. The short-term magnitudes of the first few
harmonics are used as timbre descriptors. Player identifi-
cation is performed using Gaussian classifier, where each
player is modelled using a single multivariate Gaussian
distribution with full covariance matrix. Four-fold cross-
validation procedure is used to obtain player identification
2.1 Acoustic analysis
The volume dierences between the recordings are nor-
malised to ensure that each recording has the same average
energy. The signal is then segmented into short overlap-
ping signal frames. Each signal frame is multiplied by
the Hamming window function. The windowed frames
are then zero padded, the Fourier transform is applied, and
the absolute value taken to provide short-term magnitude
spectrum. The magnitude values are compressed by ap-
plying logarithm. We used here the short-term magnitudes
of the first four harmonics as timbre descriptors. Harmon-
ics were identified semi-automatically based on the anno-
tation, i.e., the harmonics were located by finding peaks
around the multiples of the note frequency provided by the
label file. In this initial study we explore the variations in
timbre between players at the note level. All notes were
extracted from the entire recordings and the analysis was
performed for each type of note separately. Notes corre-
sponding to ornaments are discarded due to their very short
duration. As shown by Keeler (1972) through his work
with organ pipes, attack and decay sections of notes are
harmonically less stable. Thus, in order to analyse only the
stable part of the notes, we use the sustained middle third
of each note instance. This was located based on manual
annotation of note onsets. When using large amount of
data, the onset detection and note transcription could be
performed automatically as in our recent studies (K
et al., 2014) (K
uer et al., 2014).
2.2 Modelling
An acoustic model is created for each player k and each
note n, denoted as λ
. The modelling in this paper is
performed using a single multivariate Gaussian distribu-
tion with full covariance matrix. The parameters, the mean
and the covariance matrix, of each model are estimated us-
ing the training data.
We consider the identification of players from a finite
set of players based on a given piece of recording from the
test data, containing one or more instances of a given note.
The feature extraction step, as described in Section 2.1,
provides a sequence of feature vectors. Considering that a
given piece consists of R instances of a note, we have a set
of feature vectors O={O
. Each note instance is repre-
sented by a sequence of feature vectors O
, . . . , o
where T
is the number of frames in the instance i and o
is an N dimensional feature vector for frame t. The overall
probability of the feature set O is calculated on model of
each player k for the corresponding note n as
) =
) =
) (1)
and the recognised player k
is found as
= arg max
). (2)
3.1 Data description
The recordings chosen for analysis are part of a corpus
of flute melodies, selected from commercially available
sources and assembled under an AHRC Transforming Mu-
sicology project (K
uer et al., 2014). They feature the
solo flute playing of Harry Bradley, Matt Molloy, Conal
O’Grada, S
eamus Tansey and Michael Tubridy who are
prominent musicians in Irish traditional music. From the
available solo unaccompanied recordings, four traditional
tunes were chosen for each of the five players (note that the
tunes are not the same across the players). The 20 tunes
vary in length, from 17 to 41 seconds, and each tune con-
tains between 147 and 311 events, including notes, orna-
ments and breaths.
3.2 Manual annotation
The original recordings, sampled at 44.1kHz with 16 bits
and stereo, were converted by summing the channels to
mono audio. Manual annotation of the recordings was per-
formed by an experienced Irish flute player using Sonic Vi-
sualiser (Cannam et al., 2010), along with the Aubio vamp
plugins Pitch Detector and Note Detector (Brossier, 2006).
The annotation provided the start and end times of each
event, the type of event (note, ornament, breath) and the F
fundamental frequency (K
uer et al., 2014).
3.3 Experimental setup
The audio signal was analysed using frames of 1024 sam-
ples with 256 samples shift between adjacent frames. The
windowed frames were zero padded to 2048.
We employed 4-fold cross-validation to obtain more sta-
tistically reliable assessment of the performance. For each
fold, three recordings from each player were combined to
estimate the model for each player and note. The remain-
ing one recording from each player was used for testing.
This training and testing was repeated four times, taking a
dierent subset as the test set each time. Overall test per-
formance is aggregated over the four folds.
3.4 Harmonic analysis
We first present a visual assessment of the relationship be-
tween magnitudes of harmonics. For each instance of a
note in each recording, the mean magnitude value of each
of the four harmonics over the sustained middle third of the
note instance signal frames, as described in Section 2.1, is
calculated. Figure 1 depicts two dimensional scatter plots
Figure 1: Scatter plots of the short-term magnitude values (in dB) at the harmonic peaks F
and F
for note E
and note A
(right). Individual players are indicated by dierent colour and tunes by dierent shape markers. Clusters
belonging to each of the five players are marked by a two standard deviation ellipse for visual reference.
of the mean magnitudes of one harmonic component (F
against other harmonic component (F
) for all instances of
note E
and A
from all recordings. In the figure, individ-
ual players are indicated by dierent colour markers and
four tunes belonging to each player are indicated by dif-
ferent shape markers. For an easy visual reference, clus-
ters belonging to each five players are marked by an el-
lipse showing two standard deviations with corresponding
colour code. The clusters depicted in Figure 1 indicate that
Player 3 (cyan) often exhibits similar levels at F
and F
whereas Player 2 (red) has a consistent F
with a variance
in F
. Player 4 (magenta) often exhibits F
magnitudes that
are higher than F
. These trends are evident across notes in
the first octave (D
to B
). We observed that in the second
octave these trends are much less obvious with the excep-
tion of F#
3.5 Identification results
This section presents results of player identification. Ta-
ble 1 shows the overall player identification accuracy when
using data from individual notes. Results were obtained
based on using single, two and three instances of notes.
The accuracy was computed as the percentage of correctly
identified instances out of the total number of instances
over five players and over four folds. It can be seen that
the recognition accuracy usually improves slightly as the
number of used note instances increases. There are few ex-
ceptions to this, mainly in the case of using 3 note instances
(i.e., the last column in the table). The decreased accuracy
in these cases may be due to a note instance that is being
modelled poorly by the correct player model aecting neg-
atively the recognition outcome in more cases than it would
have if basing the recognition on single note instances. It
can be seen that the highest accuracy is achieved for note
which is 57%, 63% and 68% respectively when using
single, two and three note instances. The second best re-
sults with the accuracy of 51%, 54% and 55% are achieved
for note A
. On the other side, using data from note F#
achieved the lowest accuracy, 19%, 22% and 26% which
is at the border of or only little above the random guess.
Player Identification Accuracy (%)
Note Number of note instances
1 2 3
35.8 38.0 34.5
57.1 62.9 67.7
18.8 22.2 26.4
41.9 40.0 37.5
50.8 53.7 55.2
44.1 49.6 48.8
36.4 36.7 37.3
40.6 40.3 39.2
47.0 49.7 46.7
Table 1: Player identification accuracy (%) for each note
when using data of one, two or three note instances.
Figure 2 shows the confusion matrices obtained when
using the best performing notes E
and A
. These results
are obtained for single note instance and counts are accu-
mulated across all 4 folds. In the confusion matrix, each
row refers to the actual player and column to the predicted
player. It can be seen that in the case of E
, there are only
few note instances for Player 1 and Player 4 which may
have artificially positively aected the overall results for
this note. In the case of A
, the best identification perfor-
mance is achieved for Player 3 and Player 4.
Table 2 shows identification accuracy across the 5 play-
ers when using all types of notes. It can be seen that Play-
ers 1, 3 and 4 are identified with considerably higher accu-
racy than Players 2 and 5.
! " # $ % ! " # $ %
! ! ! & & " ! "# !# ' & &
" & $ & & !( " "( $% " & ")
# & & "& & !$ # ( # &' !& !"
$ & & " ! & $ & ! ) %$ #
% & !" ! & %( % & ## #& & "%
Figure 2: Confusion matrices for identification of players, using single note instance, for notes E
and A
. The rows refer
to the actual player and the columns to the predicted player and the values represent counts.
Player Identification Accuracy (%)
Number of note instances
1 2 3
Player 1 51.1 57.9 58.5
Player 2 31.3 30.1 28.1
Player 3 49.1 53.4 55.4
Player 4 56.7 53.8 50.6
Player 5 32.6 35.3 36.3
Average 44.2 46.1 45.8
Table 2: Player identification accuracy (%) obtained for
each player when using data from all notes and one, two or
three note instances.
This paper presented our initial work on identification of
professional players in Irish traditional flute music using
short-term spectral magnitudes of harmonics as timbral de-
scriptors and Gaussian classifier.
It was shown that the timbral variations between pro-
fessional players are small but often show dierences that
are identifiable across a range of dierent notes. Players
at this level often show F
harmonics at the same level or
in excess of F
. There are variations between harmonic
levels on individual notes but at this level of playing tim-
bre depends on the magnitudes of the higher harmonics
in comparison to the fundamental (Fletcher, 1975; Pollard,
The player identification performance of the current sys-
tem is rather low, which may be due to several factors as
summarised below. The use of a single Gaussian distribu-
tion for modelling may not be sucient. A more complex
model, e.g., Gaussian mixture model, will be employed
in our future work. The amount of training and testing
data for some notes was very small. This is due to the
recordings being in a range of keys which has a bearing
on the occurrences of certain notes within the tunes. The
audio we used here is from commercial recordings. This
has an advantage of being available in the public domain,
however, the disadvantage is that there is little specific in-
formation with regard to room acoustics, choice of micro-
phone or equalisation added during the recording and mix-
ing process. In addition, the players may have used dif-
ferent makes of flute instruments. In future work we will
use a larger dataset that incorporates multiple repetitions
of each tune played by the same player to minimise the
eect of dierent instruments and recording techniques.
We will also study an employment of compensation tech-
niques, working either in the feature or model domain, to
compensate for these eects.
Recreating the tests with control over the recording pro-
cess and selection of corpus would result in a study show-
ing a more realistic comparison between players and allow
for deeper study into individual players’ timbral treatment
of specific parts of tunes as well as discovering whether
consistency exists across a variety of dierent tune types.
Looking in detail at individual notes in line with Schouten
(1968)’s acoustic parameters would also extend this re-
search. In particular, analysis of attack phases and changes
in spectral envelope during notes would be of interest.
This work was partly supported by the Arts and Human-
ities Research Council (AHRC) under the Transforming
Musicology programme.
Ali-MacLachlan, I., K
uer, M., Jan
c, P., Williams, I., &
Athwal, C. (2013). Quantifying Timbral Variations in
Traditional Irish Flute Playing. In Proc. of the Third
Int. Workshop on Folk Music Analysis (FMA), Amster-
dam, Netherlands.
Breathnach, B. (1996). Folk music and dances of Ireland. Lon-
don, UK: Ossian.
Brossier, P. M. (2006). Automatic annotation of musical au-
dio for interactive applications. PhD thesis, Queen Mary,
University of London.
Burred, J. J., Robel, A., & Sikora, T. (2010). Dynamic spectral
envelope modeling for timbre analysis of musical instru-
ment sounds. IEEE Tran. on Audio, Speech, and Lan-
guage Processing, 18(3), 663–674.
Cannam, C., Landone, C., & Sandler, M. (2010). Sonic visu-
aliser: An open source application for viewing, analysing,
and annotating music audio files. In Proc. of the Int. Conf.
on Multimedia, (pp. 1467–1468). ACM.
Chudy, M. & Dixon, S. (2010). Towards Music Performer
Recognition Using Timbre. In Proc. of the 3rd Int. Conf.
of Students of Systematic Musicology, Cambridge, UK.
Coltman, J. W. (1968). Sounding mechanism of the flute and
organ pipe. The Journal of the Acoustical Society of Amer-
ica, 44(4), 983–992.
Erickson, R. (1975). Sound Structure in Music. University of
California Press.
Fletcher, N. H. (1975). Acoustical correlates of flute perfor-
mance technique. The Journal of the Acoustical Society
of America, 57(1), 233–237.
Hamilton, S. C. (1990). The Irish Flute Player’s Handbook.
Breac Publications, Eire.
Handel, S. (1995). Timbre Perception and Auditory Object
Identification. In B. C. J. Moore (Ed.), Hearing. Aca-
demic Press.
Jensen, K. (1999). Timbre models of musical sounds. PhD
thesis, Department of Computer Science, University of
Keeler, J. (1972). The attack transients of some organ pipes.
IEEE Trans. on Audio and Electroacoustics, 20(5), 378–
uer, M., Ali-MacLachlan, I., Jan
c, P., & Athwal, C.
(2014). Automated detection of single-note ornaments
in Irish traditional flute playing. In Int. Workshop on Folk
Music Analysis (FMA), Istanbul, Turkey.
uer, M., Jan
c, P., Ali-MacLachlan, I., & Athwal, C.
(2014). Automatic detection of single and multi-note or-
naments in Irish traditional flute playing. In Proc. of the
15th Int. Society for Music Information Retrieval Conf.
(ISMIR), (pp. 15–20)., Taipei, Taiwan.
uer, M., Kearney, D., Ali-MacLachlan, I., Jan
c, P., &
Athwal, C. (2014). Towards the creation of digital library
content to study aspects of style in Irish traditional music.
In Proc. of the Int. Workshop on Digital Libraries for
Musicology, London, UK.
O’Neill, F. (1998). O’Neill’s music of Ireland. Pacific, Mis-
souri: Mel Bay Publications.
Pollard, H. (1996). Timbre and Loudness of Flute Notes. Acous-
tics Australia, 24, 45–46.
Schouten, J. F. (1968). The perception of timbre. In Reports of
the 6th Int. Congress on Acoustics, volume 76, (pp.
Timoney, J., MacManus, L., Lysaght, T., & Schwarzbacher, A.
(2004). Timbral attributes for objective quality assessment
of the Irish tin whistle. In Proc. of the 7th Int. Conf. on
Digital Audio Eects (DAFX), volume 1, Perugia, Italy.
Widholm, G., Linortner, R., Kausel, W., & Bertsch, M. (2001).
Silver, gold, platinum - and the sound of the flute. Proc.
of the Int. Symposium on Musical Acoustics, 1, 277–280.
Williams, S. (2010). Irish Traditional Music. Abingdon, Oxon:
... The presented digital library is aimed to serve users in both the musicology and technology communities, from musicians to academics in this field to develop an understanding of style through the analysis of performances of a core repertoire by a variety of musicians. We have used a part of the library in our previous research [3,4,24,36]. This paper starts by introducing in Sect. 2 the structure of Irish tunes, the types of tunes and ornaments. ...
... In all of the forty-four tracks recorded by McKenna, he is accompanied by other musicians. 3 For the purposes of the archive, tracks 14. Bernard Flatherty that feature just McKenna and a piano accompanist (who is merely playing or 'vamping' chords) have been selected for future inclusion. ...
Full-text available
This paper presents the curation and annotation of a collection of traditional Irish flute recordings to facilitate the analysis of stylistic characteristics. We introduce the structure of Irish tunes, types of tunes and the ornamentation, which is a decisive stylistic determinant in Irish traditional music. We identify seminal recordings of prominent flute players and provide information related to players and their style and geographical context. We describe the process of manual annotation of the audio data. The annotations consist of the onsets of notes, note frequency and identity of notes and ornaments. We also present initial stylistic analysis of individual players in terms of ornamentation and phrasing and provide a variety of statistics for the data. The ability to accurately represent and analyse stylistic features such as ornaments allow for the development of discourse related to several key ethnomusicological questions surrounding music making, musical heritage and cultural change.
Focus: Irish Traditional Music is an introduction to the instrumental and vocal traditions of the Republic of Ireland and Northern Ireland, as well as Irish music in the context of the Irish diaspora. Ireland's size relative to Britain or to the mainland of Europe is small, yet its impact on musical traditions beyond its shores has been significant, from the performance of jigs and reels in pub sessions as far-flung as Japan and Cape Town, to the worldwide phenomenon of Riverdance. Focus: Irish Traditional Music interweaves dance, film, language, history, and other interdisciplinary features of Ireland and its diaspora. The accompanying CD presents both traditional and contemporary sounds of Irish music at home and abroad.
Conference Paper
This paper presents an automatic system for the detection of single- and multi-note ornaments in Irish traditional flute playing. This is a challenging problem because ornaments are notes of a very short duration. The presented ornament detection system is based on first detecting onsets and then exploiting the knowledge of musical ornamentation. We employed onset detection methods based on signal envelope and fundamental frequency and customised their parameters to the detection of soft onsets of possibly short duration. Single-note ornaments are detected based on the duration and pitch of segments, determined by adjacent onsets. Multi-note ornaments are detected based on analysing the sequence of segments. Experimental evaluations are performed on monophonic flute recordings from Grey Larsen’s CD, which was manually annotated by an experienced flute player. The onset and single- and multinote ornament detection performance is presented in terms of the precision, recall and F-measure.
Spectrumanalysisof flute soundspublishedby Fletcher [J.Acoust.Soc.Am.57,233-237(1975)] has been used to computeloudnessleveland tristimuluscoordinatesfor three notes C4, Cs, C6 playedboth loud and soft by four players. The differingtimbre valuesfor the same note played by the four flutists and the differences betweenloud and soft nates are clearlyreveaiedin tristimulusdiagrams.
Conference Paper
This paper presents our initial work on collection of recordings and related metadata with a view to the creation of digital library content for analysis of stylistic characteristics in Irish traditional music. We focus on ornamentation as this is a decisive stylistic determinant in Irish traditional music. The digital library contains a collection of audio recordings of prominent Irish flute players and metadata related to these recordings, such as manual annotation of onsets and offsets, identity of notes and ornaments, information related to performers and performers' style and make and type of instrument used.