Freely available online

ENGINEERING REPORTS
D. Ackermann, F. Brinkmann, and S. Weinzierl,
“A Database with Directivities of Musical Instruments,”
J. Audio Eng. Soc., vol. 72, no. 3, pp. 170–179 (2024 Mar.).
https://doi.org/10.17743/jaes.2022.0128.

A Database with Directivities of Musical

Instruments

DAVID ACKERMANN,∗ FABIAN BRINKMANN, AES Associate Member AND STEFAN WEINZIERL
([email protected]) ([email protected]) ([email protected])

Audio Communication Group, Technische Universität Berlin, Germany

This article presents a database of recordings and radiation patterns of individual notes for 41
modern and historical musical instruments, measured with a 32-channel spherical microphone
array in anechoic conditions. In addition, directivities averaged in one-third-octave bands
have been calculated for each instrument, which are suitable for use in acoustic simulation
and auralization. The data are provided in Spatially Oriented Format for Acoustics. Spatial
upsampling of the directivities was performed based on spherical spline interpolation and
converted to OpenDAFF and Generic Loudspeaker Library formats for use in room acoustic
and electro-acoustic simulation software. For this purpose, a method is presented for how
these directivities can be referenced to a specific microphone position in order to achieve a
physically correct auralization without coloration. The data is available under the CC BY-NC
4.0 license.

0 INTRODUCTION either octave or one-third-octave band averaging. The di-

rectivity data itself, however, was not made accessible with
Studies of the sound radiation characteristics of the hu- the two published studies.
man voice date back to the late 1930s [1], and studies of In a more recent approach, the radiation patterns of 14
the directivity of musical instruments began 30 years later musical instruments was measured with a combination of
(summarized in [2, Chap. 4]). While these early measure- single-capture recordings with an arc-shaped microphone
ments were often made with a single microphone moved array and repeated measurements with musicians rotated
around the source, the radiation patterns of acoustic sound stepwise under this array, providing a total of 2,522 sam-
sources such as loudspeakers, singers, or musical instru- pling points on a sphere. The inherent problem of the re-
ments are now usually measured with the source at the peatability of the notes played was partially corrected by
center of an enclosing microphone array in anechoic con- using the spectral distribution of the notes at a reference
ditions. In the following, the authors give an overview of microphone, which allowed the correction of deviations
existing databases on the directivity of musical instruments. in timbre but not those caused by the movements of the
An overview of the directivity of the human voice is beyond musicians. These directivities are freely available in one-
the scope of this article, and interested readers are referred third-octave resolution [7], while the recorded individual
to the reviews by Abe [3] and Pörschmann [4]. notes are not published as such.
The directivity of a selection of eight orchestral instru- One of the most comprehensive public databases to date
ments was measured acoustically using a 64-channel spher- includes 41 contemporary and historical instruments, in-
ical microphone array [5]. For each instrument, the direc- cluding a soprano. This resource used 32 microphones
tivity of 12 fundamental tones and their associated six to 16 for its measurements and features recordings of individ-
harmonics were quantified. Using correlation analysis, the ual notes within the playable range of each instrument and
radiation patterns of the derived partials were analyzed for directivities derived from the stationary portions of these
each instrument to determine their interdependence. In a notes [8, 9].
similar way, 14 instruments of a symphony orchestra were Based on these measurements,1 the aim of this work was
measured using a 22-channel microphone array [6], and to provide the directivities of musical instruments in an
the directivity of each instrument was characterized using

1
*To whom correspondence should be addressed: e-mail: The data are taken from version 2 of the repository:
[email protected] https://doi.org/10.14279/depositonce-5861.2.

170 J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar.

ENGINEERING REPORTS DIRECTIVITY OF MUSICAL INSTRUMENTS

open and standardized format, including both the acous- Table 1. List of modern (m) and historical (h) instruments
tic measurements, the results of the subsequent process- whose individual directivities are available in the database in
the indicated pitch range for the playing dynamics pianissimo
ing, and important metadata, such as the exact position (pp) and fortissimo (ff).
of the microphone capsules, tuning frequency of the in-
strument, and pitch for which the directivity is valid. This Scale
should facilitate the exchange and use of such data by the
scientific and general acoustics community. To facilitate Group Instrument Era pp ff
this, the recently standardized Spatially Oriented Format
Strings Violin m G3–G6 G3–G6
for Acoustics (SOFA) convention FreeFieldDirec- Violin h G3–C7 G3–C7
tivityTF was used [10, 11], but the data are also pro- Viola m C3–F7 C3–F7
vided in OpenDAFF [12] and Generic Loudspeaker Li- Viola h C3–C6 C3–C6
brary (GLL) [13] formats for use in room acoustic simu- Cello m C2–G6 C2–G6
lation software. For each instrument, the database contains Cello h C2–D5 C2–D5
Double bass m E1–E5 E1–D#5
the single-note recordings [calibrated to an absolute sound Double bass h E1–E3 E1–E3
pressure and equalized for the microphone array transfer Acoustic guitar m C2–B5 E2–B5
function (TF)], extracted single-tone directivities for each Double action harp m C1–F4 C1–F4
partial tone, and one-third-octave band averaged directivi- Woodwinds Oboe m A#3–A6 A#3–F#6
ties and corresponding finite impulse responses (FIRs) [9]. Classic oboe h C4–D#6 C4–D#6
Romantic oboe h B3–C6 D4–G5
English horn m E3–G#5 E3–G#5
Clarinet m D3–A#6 D3–A#6
1 METHODS Clarinet h D3–F6 D3–F6
Bass clarinet m A#1–D#5 A#1–D#5
The directivity of 40 modern and historical string, wood- Bassoon m A#1–E5 A#1–E5
wind, brass, and percussion instruments and one soprano Classic bassoon h A#1–C5 A#1–C5
singer was measured using a 32-channel full-spherical mi- Baroque bassoon h A1–F#4 A1–F#4
crophone array in an anechoic chamber. Table 1 gives an Contrabassoon m A#0–D#3 A#0–F3
overview of the instruments contained in the data base in- Dulcian h C2–G4 C2–G4
Alto saxophone m C#2–C#5 C#2–C#5
cluding the range of musical notes for which recordings are Tenor saxophone m B1–A4 B1–A4
available. The measurements were presented in an earlier Flute m B3–A#6 B3–D7
publication [8]. The following sections detail the record- Transverse flute h D4–F#6 D4–A#6
ing of the original database and highlight the extended and Keyed flute h C#4–A6 C#4–A6
improved processing of the data in this work. Brass Trumpet m F#3–F6 F#3–F6
Natural trumpet h D3–D6 D3–D6
French horn m D2–F#5 D2–G5
1.1 Data Representation and Format Natural horn h A1–B4 A1–B4
Basset horn h F2–A#5 F2–A#5
The directivity of electro-acoustic sound sources, such as Tenor trombone m G1–F5 G1–F5
loudspeakers and microphones, can be described relatively Bass trombone m E1–F4 E1–F4
easily by using TFs in the frequency domain or by FIR Alto trombone h D2–D#5 C#2–D#5
filters in the time domain for each direction of radiation. Bass trombone m B1–B4 C1–B4
Tenor trombone h E1–D5 E1–D5
In contrast the directivity of natural sound sources, such as Tuba m G#0–C5 F0–E5
human speakers, singers and musical instruments, is more Percussion Timpani m D2 D2
complex to describe, because the directivity of a musical Pedal timpani m D2 D2
instrument depends not only on the frequency but also on Soprano ··· G3–G#5 G3–G#5
the note being played, and sometimes on the fingering, so
that the same note can have multiple radiation patterns. The
different openings for the sound radiation of a bassoon, octave band averaged data (see SEC. 1.3.3). This represen-
which can be open or closed depending on the note played tation is mainly used in geometric acoustic simulation.
and fingering used, are shown in Fig. 1 as an example. The For a complete description of natural sound source direc-
note-dependent directivity has also been shown to produce tivities, information about the instrument/singer and way of
audible differences in sound [14]. playing is needed in addition to the measurement setup. The
To ensure maximum flexibility, the Freefield Di- SOFA standard allows these data to be stored as metadata
rectivityTF convention represents directivity data as and uniquely assigned for easy handling and data exchange.
complex TFs at arbitrary frequencies. This allows FIR fil- An overview of the available metadata can be found in SEC.
ters of artificial sound sources but also multi-channel acous- 2.
tic recordings of a musical instrument (see SEC. 1.3.1) to
be stored as complex spectra with linear frequency reso- 1.2 Measurement Setup
lution. In addition the directivity patterns of natural sound The instruments were recorded with a fully spherical
sources can be stored for arbitrary, not necessarily equidis- lightweight microphone array in the anechoic chamber of
tant, frequencies such as the fundamental frequency and the Technische Universität Berlin with a room volume of
corresponding overtones (see SEC. 1.3.2), or as one-third- approximately 1,070 m3 and lower cut-off frequency of fc =

J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar. 171

ACKERMANN ET AL. ENGINEERING REPORTS

Fig. 1. The sound radiation of the bassoon for the note E#3 in five different frequency bands, from 1 to 5 kHz, recorded simultaneously
with an acoustic camera. Bright spots indicate high sound radiation from the instrument. Distributed, frequency-dependent sound
radiation becomes visible from the open tone holes on the body, bell, and mouthpiece. Depending on the note played and fingering used,
these tone holes can be open or closed, leading to different sound radiation, even for the same frequency [15].

63 Hz. Thirty-two Sennheiser KE4-211-2 electret capsules DirectivityTF convention requires frequency data, the
of the microphone array were located at the vertices of a recordings were Fourier transformed
pentakis dodecahedron with a radius of r = 2.1 m, a central N −1
angle of  = 37.4◦ , and an arc length of s = 1.37 m between 
xq [n]e−i2π N n ,
k
X q (k) = (1)
the individual microphones.
n=0
A height-adjustable chair was used to position the musi-
cians with their instruments so that the estimated acoustic with i2 = −1 being the imaginary unit, and saved in the
center of each instrument was as close as possible to the SOFA files as complex-valued single-sided spectra XS, q (k)
center of the microphone array, and the musicians faced of length N/2 + 1
the positive x-axis. For extended instruments with sec- ⎧
⎨ X q (k), if k = 0
ondary sound sources, such as the bassoon with its bell
X S,q (k) = 2 · X q (k), if 0 < k < N2 (2)
and multiple spaced tone holes, the geometric center of the ⎩
X q (k), if k = N2 .
sound-emitting surface of the instrument has been defined
as the acoustic center. The musicians were asked to play Note that reconstructing the time domain recordings from
all playable notes of their instruments in both extremes of the published data thus requires the reconstruction of the
the dynamic strength, i.e., in pianissimo (pp) and fortissimo both-sided spectrum of length N
(ff) (for precise instructions see [16]), without moving the ⎧
instrument; the instruments were not fixed by any technical ⎪
⎪X S,q (k), if k = 0
⎨ 1
· X (k), if 0 < k < N2
S,q
means during the recording. X q (k) = 2 , (3)
The exact capsule positions of the array are given in ⎪
⎪ X S,q (k), if k = N2
⎩1 ∗
spherical coordinates, i.e., in azimuth (φ = 0◦ point- 2
· X S,q (N − k), if k > N2
ing in positive x-direction, φ = 90◦ pointing in positive with ( · )* denoting the complex conjugate, before applying
y-direction), colatitude (θ = 0◦ pointing in positive x- the inverse Fourier transform (see [17] for details).
direction, θ = 90◦ pointing in positive z-direction), and
distance (in meters) are included in the metadata of the
1.3.2 Single Tone Directivity Data
SOFA files (cf. SEC. 2). The recordings of the single notes
of each instrument were made with 24-bit resolution and a To determine the directivity of the musical instruments,
sampling frequency of fs = 44.1 kHz. The recordings were the authors used the stationary part of the single note record-
calibrated—i.e., a digital amplitude of 1 corresponds to a ings. For all instruments producing stationary parts, this was
pressure of 1 Pascal or a sound pressure level of Lp = 94 dB manually windowed by visual inspection, resulting in dura-
– and compensated for the frequency response of the mi- tions between 200 and 2,104 ms, with a median duration of
crophone array and capsules. The exact measurement setup 630 ms. For the acoustic guitar and harp, a quasi-stationary
and a detailed description of the calibration procedure can part was defined as the part between the decay time and
be found in [8]. release time as estimated with the Timbre Toolbox [18].
For the transient timpani signals, the entire recording was
used, from the onset to the transition into the noise floor.
1.3 Processing The directivities were estimated in two steps. First the
The calibrated and equalized single-note recordings in ff fundamental frequency f0 and frequency of the overtones fi
and pp provided as 32-channel WAV files, as published in in hertz were identified with i ∈ {1, 2, ..., I} and I being
[9], form the basis for the processing described below. the highest identifiable overtone. Secondly the energy at
these frequencies was estimated. Both steps were based on
the magnitude response |Xq (k)| and ignored the phase infor-
1.3.1 Recordings mation. Determining the phase of a directivity TF would
The real-valued single-note recordings xq [n] of even require a reference signal. While for loudspeaker directiv-
length N are available at discrete times n ∈ {0, 1, ..., N ities the input signal can be used as a reference, there is
− 1} and for Q = 32 channels of the spherical microphone no useful reference for a musical instrument, because it is
array with q ∈ {1, 2, ..., Q}. Because the FreeField- an extended sound source with a note-dependent acoustical

172 J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar.

ENGINEERING REPORTS DIRECTIVITY OF MUSICAL INSTRUMENTS

for physical reasons, the partials do not always lie exactly at

the harmonic multiples of the fundamental frequency and
that the fundamental was not always held exactly constant
over the duration of the stationary part. If the identified
frequency deviated from the harmonic frequencies by more
than 5 cents, the procedure was stopped, and the signal en-
ergy was considered to be below the noise floor. A visual
inspection of the detected signal components confirmed
that all relevant and clearly identifiable partial tones had
been found by this procedure.
At the frequencies fi that could be estimated, the directiv-
ity was calculated using the power spectral density (PSD):

1
Sx x,q (k) = |X q (k)|2 , (4)
fs N
which is a measure of the power of each frequency compo-
nent in a signal with the implicit unit Pa2 /Hz. The authors
Fig. 2. Extract from the spectrogram for the signal of a trumpet
playing note A4 (f0 = 440 Hz) at fortissimo (microphone 4), with have used the Welch method to improve the robustness of
amplitude in decibels (a) and phase in degree (b). the PSD against noise by

1) Dividing the signal into eight segments of equal

center and small but certainly phase-relevant movements length with 50% overlap,
of the player. The non-stationary acoustic center is evident, 2) Applying a Hanning window function to each seg-
for example, in deviations of the phase response in Fig. 2 ment to reduce the effect of spectral leakage caused
from a strict periodicity that would be expected at constant by the finite length of the segments,
travel time. Any reference, and thus any phase representa- 3) Computing the periodogram of each segment, which
tion in the directivity of a musical instrument, would thus is an estimate of the PSD of that segment, and
be arbitrary. 4) Averaging the periodograms across the segments to
With regard to the use of directivity for room acoustic obtain an estimate of the overall PSD of the signal.
simulation and auralization, experience in physical rooms
shows that small changes in the position of a sound source in To obtain an estimate of the power at each frequency,
space, although leading to substantial changes in the phase each estimate of the PSD was scaled by the equivalent
response of the TF, do not produce notable differences in noise bandwidth of the window (in hertz). Finally the power
sound, at least for natural, acoustic instruments. Thus, even was determined by simple peak picking in the scaled PSD
for such applications, the phase can be chosen essentially around the frequencies fi of the tone’s harmonics and con-
freely. verted to the sound pressure value pq (fi ) by taking the square
Finally also the interpolation of the phase spectrum is root. Fig. 3 illustrates the process for one note played by
highly susceptible to noise and can lead to errors, espe- the modern oboe.
cially at high frequencies [19]. The authors have therefore
proposed the use of absolute-valued directivities for inter- 1.3.3 One-Third-Octave Band Averaging
polation [20]. This is in contrast to the previous processing Geometric acoustics applications typically require a sin-
of the data, which used the complex valued spectrum [8]. gle directivity pattern for a sound source, which is usually
An estimate of the fundamental frequency f0 was made provided in one-third-octave band resolution. This was cal-
by identifying the frequency with the highest amplitude culated by energetic averaging of the partials of all individ-
within a window of ±100-cent bandwidth around the fre- ual notes of an instrument falling into the one-third-octave
quency corresponding to the tuning pitch indicated by the bands according to IEC 61260-1 [21], for the M = 30 cen-
musicians. This was usually 442 or 443 Hz for modern in- ter frequencies from 25 Hz to 20 kHz. All partials of the J
struments, 430 Hz for instruments of the Classical period, individual notes pq (fi,j ) from SEC. 1.3.2 were used for the
and 415 Hz for instruments of the Baroque period. This one-third-octave band representation. The data were av-
frequency was obtained for all 32 microphone recordings, eraged for each of the Q = 32 receivers of the spherical
and the most frequently occurring frequency over all 32 microphone array individually, by calculating the averaged
extracted values was chosen as f0 . amplitude as
In the next step, the frequencies of the partials were es- 
 L−1
timated by placing a search window of ±10 cents around 1 
each harmonic frequency corresponding to f0 and identify- p̄q,m = p2 , (5)
L i=0 q,i
ing the most frequently occurring, highest amplitude within
this window, again considering all 32 microphone record- where L indicates the number of partials identified in one
ings. The search window takes into account the fact that, one-third-octave band. Fig. 4 illustrates the averaging pro-

J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar. 173

ACKERMANN ET AL. ENGINEERING REPORTS

Fig. 3. Normalized FFT (gray) and scaled PSD (black) of the signal of the modern oboe playing note C4 (262 Hz) at fortissimo, recorded
with microphone 4. The asterisks indicate the detected power values and frequencies f0 and fi that do not necessarily coincide with the
discrete bins of the PSD. SPL = sound pressure level.

where wq are the normalized area weights of the measure-

ment grid with wq = 1. This representation of the radi-
ation patterns is called diffuse equalization in the following
and will be discussed in more detail in SEC. 3.
The final step was to calibrate the directivity to the sound
power of the real instruments, because the previous diffuse
equalization [cf. Eq. (6)] had lost the absolute sound power
reference. This was done by averaging the sound pressure
level over the effective one-third-octave bands for each mi-
crophone. The one-third-octave bands with no sound radi-
Fig. 4. Estimated sound pressure levels (SPLs) of all partials of ation were set to zero.
the modern oboe plotted against frequency (gray) in the one-third-
The average sound pressure level of the diffuse equalized
octave bands from 400 to 2.5 kHz, recorded with microphone 13.
As an example, the fundamental and the first five overtones for directivity per microphone was then calculated as
the note A4 (f0 = 442 Hz) are highlighted in black. The dashed ⎛ 1 ⎞
2
vertical lines mark the boundaries of the one-third-octave bands. M
p̄diff,q,m
The horizontal solid black lines indicate the energetically averaged ⎜ m∈M ⎟
L P,3rd,q = 10 lg ⎝ ⎠, (7)
sound pressures p̄q,m according to Eq. (5), shown in decibels. p02

cedure over all partials for the modern oboe in one-third- where M indicates the number of effective one-third-octave
octave bands from 400 to 2.5 kHz. bands and p0 = 2 × 10−5 Pa. Its average over the surface
At this point, the data still contain the direction- of the spherical envelope is given by
dependent frequency response of each instrument. If the ⎛ ⎞
directivities are used for auralizations in which a simu- 1 Q

lated (binaural) room impulse response is convolved with L̄ P,3rd = 10 lg ⎝ 100.1L P,3rd,q ⎠ . (8)
Q q=1
an anechoic recording of an instrument, it should be noted
that the anechoic recording also contains the frequency re-
The sound pressure of the reference, i.e., the corresponding
sponse of the instrument in the direction of the recording
instruments, was calculated from the calibrated recordings
microphone. To avoid an unnatural coloration during au-
as follows:
ralizations, the frequency response has to be removed from
⎛ ⎞
the one-third-octave band averaged directivities. In theory 1
N J
2
this normalization should also consider the position of the ⎜ N J n=1 j=1 xq, j [n] ⎟
⎜ ⎟
microphone from which the anechoic recording of the in- L P,ref,q = 10 lg ⎜ ⎟, (9)
⎝ p02 ⎠
strument was made, i.e., this direction should be normalized
to 0 dB. If this position, however, is unknown or unstable
due to movements of the instrument relative to the micro-
where N is the number of samples of the stationary part
phone, it may be most robust to equalize the directivity so
and J is the number of single notes and averaged over all
that equal energy is radiated across all directions within
microphones to
each band, i.e., to
⎛ ⎞
p̄q,m 1 Q
p̄diff,q,m =
Q
, (6) L̄ P,ref = 10 lg ⎝ 100.1L P,ref,q ⎠ . (10)
2
p̄q,m · wq Q q=1
q=1

174 J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar.

ENGINEERING REPORTS DIRECTIVITY OF MUSICAL INSTRUMENTS

Finally, the calibration of the diffuse equalized directivity

from Eq. (6) is given by
L̄ P,ref − L̄ P,3rd
p̄cal,q,m = p̄diff,q,m · 10 20 . (11)
An alternative way to calculate the one-third-octave band
averaged directivity would have been to stitch all single-
note recordings, use band filters, and analyze the complete
signal within each one-third octave, which would result
in an energy-weighted average of the individual frequency
components. Due to the strong predominance of the tonal Fig. 5. The one-third-octave band averaged radiation pattern of
parts (see Fig. 3), this would, however, yield the same results the modern bassoon at 400 Hz, with original measurement resolu-
compared to the authors’ method within a range of precision tion (32 points, a) and spatial interpolation to 2,522 points, using
that is reasonable for directivity measurements. The one- spherical splines of first order (right). Balloon representation with
third-octave band averaged directivity of each instrument radius corresponding to the sound pressure level in decibels.
is calculated for the dynamic fortissimo (ff) and provided
in the FreeFieldDirectivityTF SOFA convention.

1.3.4 Spatial Interpolation

Several applications that rely on the directivity of sound
sources, such as room acoustic simulations, require the use
of continuous or high resolution data. Consequently, the
measurement data must be spatially resampled (interpo-
lated) to match the required sampling grid.
By sampling the actual sound pressure function f(θ, φ)
with a Q channel spherical microphone array, the samples
pq = f(θq , φq ) are given at the positions (θq , φq ) of the
respective microphones for q ∈ {1, 2, ..., Q}. The general
mathematical formula for interpolation can therefore be
expressed as

Q
f̂ (θr , φr ) = f (θq , φq ) · L q (θr , φr ), (12)
q=1

where f̂ (θr , φr ) = p̂r is the estimated sound pressure at the

R points (θr , φr ) of the interpolation grid for r ∈ {1, 2, ...,
R} and Lq (θr , φr ) is the interpolation function derived from
Fig. 6. The one-third-octave band spectrum (gray) and smoothed
the known sound pressure pq at the position (θq , φq ). The one-third-octave band spectrum (black) for the instrument bassoon
specific choice of the interpolation function depends on the modern (a) for the radiation direction (θ = 0; φ = −10) and
interpolation method being used. corresponding energy-time curve (b). SPL = sound pressure level.
There is a plethora of techniques for interpolating real-
valued scattered data that make different assumptions about
the distribution of the discrete set of known data points
[22]. For musical instruments, the thin-plate pseudo-spline 1.3.5 One-Third-Octave Band Band Smoothed
method [23, 24] of order 1 has been found to be a good FIRs
method, producing lower interpolation errors than spheri- To enable musical instruments to be used as sound
cal harmonics interpolation [25–27] or three-dimensional sources for simulating room acoustic and electroacoustic
Vector-Based Amplitude Panning [28] when applied to environments with software that uses FIR filters to repre-
sparsely sampled directivity measurements and evaluated sent directivity, R = 2,522 FIR filters were calculated using
against the directivity of different musical instruments mea- the above-mentioned grid. For this purpose, a one-third-
sured at high resolution as a reference [20]. octave band spectrum according to IEC 61260-1 [21] with
The authors chose an equiangular grid with an angular a frequency resolution of 1 Hz was generated from the es-
resolution of 5◦ in azimuth and colatitude as the target for timated sound pressure values p̂r,m (gray line in Fig. 6).
the interpolation, resulting in R = 2,522 sound pressure The one-sided spectrum with odd N was then smoothed
values p̂r,m for each of the m ∈ M one-third-octave bands. with a one-third-octave band filter (black line in Fig. 6).
The closed-form spherical spline interpolation for order After transformation to a two-sided spectrum according to
1 was realized with AKtools using the function AKsph- Eq. (3), the spectrum was transformed into the time do-
SplineInterp() [29] based on the directivity patterns main using inverse fast Fourier transform (IFFT), and the
according to Eq. (11). phase was made minimum phase using the AKTools func-

J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar. 175

ACKERMANN ET AL. ENGINEERING REPORTS

tion AKphaseManipulation(). Finally the FIR filter the TF in hertz. Note that the calibrated 32-channel WAV
was reduced to 8,192 samples according to AES56-2008 recordings on which this data set is based are still freely
[30] as shown in Fig. 5. accessible [9].4
For use in the software EASE,2 this data was stored in
the proprietary GLL format. 2.2 SOFA Single Note
There is also a separate SOFA file for each instru-
2 DATABASE ment and note for the single-note directivity data; the
naming of the file corresponds to the scheme Source-
The database contains the calibrated single-note record- Name midiNote singleTones.sofa. The purely
ings, single-note directivities, and frequency-averaged di- real sound pressure levels are stored as complex TFs in
rectivity patterns in SOFA format FreeFieldDirec- the field Data.Real with the dimension MxRxN, where
tivityTF convention under a Creative Commons share M is always 1, R = 32, and N refers to the number of
alike license (CC-BY-NC 4.0). In addition, high spatial the I extracted partials. The Data.Imag field with the di-
resolution–interpolated radiation patterns averaged over mension MxRxN is included in the dataset for consistency
one-third-octave bands are provided in openDaff format reasons but contains only zeros. The field N indicates the
and as FIR filters in GLL format. All data are freely acces- frequencies of the I partials in hertz.
sible [9]. A list of the available musical instruments can be
found at Table 1. 2.3 SOFA One-Third-Octave Band
The SOFA files can be read using a variety of APIs.3 For the one-third-octave band averaged directivities,
They contain the recorded signals and extracted directiv- there is one SOFA file for each instrument with the nam-
ities as complex TFs together with metadata describing ing scheme SourceName 3rdOctave.sofa. The av-
the data in detail. This includes the name of the instru- eraged and calibrated sound pressures from Eq. (11) are
ment in the entry GLOBAL SourceName, the name of the stored in Data.Real with the dimension MxRxN, where
musician in GLOBAL Musician, the manufacturer of the M is always 1, R = 32, and N = 30 refers to the nominal
instrument in GLOBAL SourceManufacturer, and a center frequencies from 25 Hz to 20 kHz according to IEC
verbal description of the position of the instrument dur- 61260-1:2014 [21]. The data field Data.Imag has been
ing the measurement in SourceView Reference. The filled with zeros. The field N indicates the center frequencies
arrangement of the capsules of the microphone array is of the one-third-octave bands in hertz. In this case, the en-
described in ReceiverPosition. tries SourceTuningFrequency, SteadyPart, and
In the SOFA data of the recordings, the re- MidiNote are not included in this data representation.
spective note and dynamic level are indicated in
GLOBAL Description, e.g., note = A4; dy- 2.4 OpenDAFF One-Third-Octave Band
namic = ff, the MIDI number in MidiNote, e.g., 69
The open source format openDAFF can be read with
for A4, and the frequency for A4 corresponds to the tuning
several APIs.5 The naming of the data follows the scheme
frequency in the entry for SourceTuningFrequency.
SourceName.daff. The directional patterns are stored
In the SOFA data of the original recordings, Steady-
with a spatial resolution of 5◦ (azimuth and colatitude),
Part indicates the range of the manually determined
i.e., each file contains one-third-octave band magnitude
stationary portion in samples. A detailed description of the
spectra at 2,522 points. These data can be used directly
structure of the database can be found in the accompanying
in the acoustic simulation environment RAVEN [31]. For
documentation.
evaluating the directivity, of both the individual notes and
frequency-averaged directivity patterns, in arbitrary spatial
2.1 SOFA Recordings resolution, the authors provide a MATLAB script as part of
The single note recordings are available as a one- the database (cf. SEC. 2.6).
sided complex TF for each instrument and note. The
data can be converted into a two-sided spectrum ac- 2.5 GLL One-Third-Octave Band FIRs
cording to Eq. (3) and converted to the time do- The proprietary GLL format allows the integration of
main by means of an IFFT. The naming of the complex sound sources into the acoustic simulation envi-
data follows the scheme SourceName dynamic midi ronment EASE. The directional patterns averaged over a
Note recordings.sofa, and the recordings are stored one-third-octave band for all 21 modern musical instru-
in the Data.Real and Data.Imag fields. The fields ments and for a soprano singer were stored as FIR filters
have the dimension MxRxN, where M (measurement) is with 8,192 taps and with a spatial resolution of 5◦ . For the
always 1; R (receiver) is the number of capsules of the mi- exact naming scheme, please refer to the documentation of
crophone array, with R = 32; and N indicates the length of the dataset [9]. The data can be converted to UNF (used
the TF. The field N contains the frequencies of the bins of

4
In version 2 of the repository: https://doi.org/
2
www.afmg.eu/en (accessed May 2, 2023). 10.14279/depositonce-5861.2.
3 5
Cf. www.sofaconventions.org/mediawiki/index.php/Software_ Cf. www.github.com/svn2github/opendaff (accessed May 2,
and_APIs (accessed May 2, 2023). 2023).

176 J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar.

ENGINEERING REPORTS DIRECTIVITY OF MUSICAL INSTRUMENTS

by Ulysses6 ) and the CLF/CIF format (used by ODEON7 For the physically correct auralization of musical instru-
and CATT-Acoustic8 ) using the proprietary SpeakerLab9 ments in virtual acoustics, frequency-averaged directivities
software from AFMG. are a compromise that has to be made for technical reasons.
Current simulation applications do not allow a straight-
2.6 Tools forward exchange of single tone representations of such
Part of the database is the Directivity demo.m directivities within a simulation run. Due to the frequency
MATLAB script that makes it possible to read the record- averaging of the input data, a tonal coloration of the simula-
ings from the SOFA data, display their spectrum graphi- tion result may occur [14]. The magnitude of this effect and
cally, and transform them into the time domain by IFFT. its influence on instrumental and room acoustic perception
This data can then be saved as a WAV file and played back will have to be determined in a subsequent study using the
with common media players. data presented.
The script also allows for the three-dimensional display When an anechoic recording of an instrument and its
of single-note and frequency-averaged directivity in the directivity is used to auralize a virtual acoustic environment,
form of balloon plots based on the SOFA data provided. it is essential to normalize the directivity of the source to
Finally the data can be evaluated at any sampling quadra- a suitable reference. In theory this would be the position
ture using spherical spline interpolation. Prerequisite for of the microphone used to make the anechoic recording.
all processing steps is the installation of AKTools10 and Because the recorded signal already contains the directional
the SOFA API for MATLAB (SOFAToolbox11 ) contained timbre characteristic of the instrument in that direction, it
therein. should not be altered by the applied directivity, as this would
result in unwanted coloration of the simulation. This means
that the directivity should be normalized, so that only a
3 CONCLUSION frequency response relative to this reference direction will
be obtained. This will be achieved by calculating
The present dataset contains recordings and radiation
patterns of the individual notes of 41 modern and historical p̂r,m
p̂pt,r,m = , (13)
musical instruments, measured with a 32-channel micro- p̂mic,m
phone array in anechoic conditions. The recordings and di-
rectivities are provided in standardized SOFA format in the where p̂mic,m is the interpolated and frequency-averaged
FreeFieldDirectivityTF convention. From these sound pressure of the M one-third-octave band in the refer-
data, averaged directivities have been calculated for each ence direction.
instrument, which are suitable for use in acoustical simula- In a real recording situation, however, the instrument be-
tion and auralization. In addition, spatially high-resolution ing played by a musician is a moving sound source. This
directivities in OpenDAFF and GLL formats have been can result in an angular displacement of the instrument
generated, allowing direct use in software such as RAVEN which can easily reach up to 47◦ when played in a standing
and EASE. position and up to 36◦ when played in a sitting position
The absolute quality of the interpolation methods used [33]. To compensate for the movement of the instrument
for this spatial upsampling obviously depends on the char- when referencing, the authors consider it useful to equal-
acteristics of the sound source, such as its acoustically effec- ize the directivity not to a point but to the average of a
tive size, the modal patterns of its sound radiating parts, and larger spherical surface area A. The distribution of orienta-
the resulting complexity of the radiation pattern. For acous- tions over this surface can be taken into account by using a
tically small sources, such as a trumpet and trombone bell or weighting function, i.e., by calculating
violin, a fairly accurate interpolation can be expected even p̂r,m
based on a measurement at 32 points. For extended sources p̂area,r,m =  , (14)
RA
with more complex radiation patterns, however, a sparse p̂r2A ,m · wr A · gr A
sampling grid may lead to increasingly poor estimates of r A =1
the far field directivity [20]. If other types of interpolation
where wr A and gr A are the normalized area and two-
prove superior in the future, such as recently investigated,
dimensional function weights, respectively, with
physically informed interpolation methods using the Eu-
ler equations as constraints [32], the interpolations applied 
RA
may be revised in the future. wr A · gr A = 1, (15)
r A =1

and rA ∈ {1, 2, ..., RA } are the RA grid points of the area A

6
www.ifbsoft.de (accessed May 2, 2023). over which the mean is to be calculated.
7
www.odeon.dk (accessed May 2, 2023). If no information about the position of the microphone
8
www.catt.se (accessed May 2, 2023). during the recording is documented, a diffuse equalized di-
9
www.afmg.eu/en/ease-speakerlab (accessed May 2, 2023).
10
Cf. www.tu.berlin/ak/forschung/publikationen/open-research- rectivity [cf. Eq. (6)] can be used to minimize the sound
tools/aktools (accessed May 2, 2023). coloration on average. For this reason, the directivity aver-
11
Cf. www.github.com/sofacoustics/SOFAtoolbox (accessed aged over one-third-octave bands has been provided in this
May 02, 2023). representation for uniformed immediate use. The extent to

J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar. 177

ACKERMANN ET AL. ENGINEERING REPORTS

which the different referencing methods of the recording [13] AFMG, “GLL (Generic Loudspeaker Library) –
position affect the perceived sound event and room acous- A New Standard for Measuring and Storing Loudspeaker
tic parameters will be clarified in a later investigation on Performance Data,” AFMG Engineering White Paper
the basis of these data. (2007 Oct.). https://www.afmg.eu/sites/default/files/2021-
07/GLL White Paper October07.pdf.
4 ACKNOWLEDGMENT [14] A. Corcuera, V. Chatziioannou, and J. Ahrens,
“Perceptual Significance of Tone-Dependent Directiv-
This work was funded by the German Research Founda- ity Patterns of Musical Instruments,” J. Audio Eng.
tion (DFG WE 4057/9-2). The authors would like to thank Soc., vol. 71, no. 5, pp. 293–302 (2023 May).
Stefan Feistel and Silke Bögelein (AFMG) for their support https://doi.org/10.17743/jaes.2022.0076.
in producing the GLL datasets. [15] M. Kob, “Impact of Excitation and Acoustic Condi-
tions on the Accuracy of Directivity Measurements,” Pro-
5 REFERENCES ceedings of theInternational Symposium on Musical Acous-
tics (ISMA), pp. 639–643 (Le Mons, France) (2014 Jul.).
[1] H. K. Dunn and D. W. Farnsworth, “Exploration of [16] S. Weinzierl, S. Lepa, F. Schultz, et al.,
Pressure Field Around the Human Head During Speech,” “Sound Power and Timbre as Cues for the Dynamic
J. Acoust. Soc. Am., vol. 10, no. 3, pp. 184–199 (1939 Jan.). Strength of Orchestral Instruments,” J. Acoust. Soc.
https://doi.org/10.1121/1.1915975. Am., vol. 144, no. 3, pp. 1347–1355 (2018 Sep.).
[2] J. Meyer, Acoustics and the Performance of Music https://doi.org/10.1121/1.5053113.
(Springer, New York, 2009). [17] J. Ahrens, C. Andersson, P. Höstmad, and W. Kropp,
[3] O. Abe, Sound Radiation of Singing Voices, Ph.D. “Tutorial on Scaling of the Discrete Fourier Transform and
thesis, Universität Hamburg, Hamburg, Germany (2019 the Implied Physical Units of the Spectra of Time-Discrete
Sep.). Signals,” presented at the 148th Convention of the Audio
[4] C. Pörschmann, “A Database for the Comparison of Engineering Society (2020 May), e–Brief 600.
Measured Datasets of Human Voice Directivity,” in Pro- [18] G. Peeters, B. L. Giordano, P. Susini, N. Misdariis,
ceedings of the Forum Acusticum , pp. 1–8 (Torino, Italy) and S. McAdams, “The Timbre Toolbox: Extracting Audio
(2023 Sep.). Descriptors From Musical Signals,” J. Audio Eng. Soc., vol.
[5] F. Hohl, Kugelmikrofonarray zur Abstrahlungsver- 130, no. 5, pp. 2902–2916 (2011 Nov.).
Messung von Musikinstrumenten, Master’s thesis, Institute [19] J. Ahrens and S. Bilbao, “Computation of Spherical
of Electronic Music and Acoustics, University of Music Harmonics Based Sound Source Directivity Models From
and Performing Arts, Graz, Austria (2009 Nov.). Sparse Measurement Data,” in Proceedings of the Forum
[6] J. Pätynen and T. Lokki, “Directivities of Sym- Acusticum, pp. 2019–2026 (Lyon, France) (2020 Dec.).
phony Orchestra Instruments,” Acta Acust. united [20] D. Ackermann, F. Brinkmann, F. Zotter, M. Kob,
Acust., vol. 96, no. 1, pp. 138–167 (2010 Jan./Feb.). and S. Weinzierl, “Comparative Evaluation of Interpola-
https://doi.org/10.3813/AAA.918265. tion Methods for the Directivity of Musical Instruments,”
[7] S. D. Bellows, K. J. Bodon, and T. W. Leishman, EURASIP J. Audio Speech Music Process., vol. 2021, paper
“Musical Instrument Directivity,” https://scholarsarchive. 36 (2021 Oct.). https://doi.org/10.1186/s13636-021-00223-
byu.edu/directivity/ (accessed Jul. 20, 2020). 6.
[8] N. R. Shabtai, G. Behler, M. Vorländer, and S. [21] IEC, “Electroacoustics - Octave-Band and
Weinzierl, “Generation and Analysis of an Acoustic Radia- Fractional-Octave-Band Filters - Part 1: Specifications,”
tion Pattern Database for Forty-One Musical Instruments,” IEC Standard 61260-1:2014 (2014 Jun.).
J. Acoust. Soc. Am., vol. 141, no. 2, pp. 1246–1256 (2017 [22] J. Li and A. D. Heap, A Review of Spatial Interpo-
Feb.). https://doi.org/10.1121/1.4976071. lation Methods for Environmental Scientists (Geoscience,
[9] S. Weinzierl, M. Vorländer, D. Ackermann, et al., Australia, Canberra, 2008).
“A Database of Anechoic Microphone Array Measure- [23] G. Wahba, “Spline Interpolation and Smoothing on
ments of Musical Instruments,” DepositOnce: Technis- the Sphere,” SIAM J. Sci. Stat. Comput., vol. 2, no. 1, pp.
che Universität Berlin, 5861, Version 3 (2024 Feb.). 5–16 (1981 Mar.). https://doi.org/10.1137/0902002.
https://doi.org/10.14279/depositonce-19858. [24] G. Wahba, “Erratum: Spline Interpolation and
[10] AES, “AES Standard for File Exchange - Spatial Smoothing on the Sphere,” SIAM J. Sci. Stat. Com-
Acoustic Data File Format,” AES Standard 69-2022 (2022 put., vol. 3, no. 3, pp. 385–386 (1982 Sep.).
Sep.). https://doi.org/10.1137/0903024.
[11] P. Majdak, F. Zotter, F. Brinkmann, et al., “Spatially [25] G. Weinreich and E. B. Arnold, “Method for
Oriented Format for Acoustics 2.1: Introduction and Recent Measuring Acoustic Radiation Fields,” J. Acoust. Soc.
Advances,” J. Audio Eng. Soc., vol. 70, no. 7, pp. 565–584 Am., vol. 68, no. 2, pp. 404–411 (1980 Aug.).
(2022 Jul.). https://doi.org/10.17743/jaes.2022.0026. https://doi.org/10.1121/1.384751.
[12] F. Wefers, “OpenDAFF: A Free, Open-Source Soft- [26] F. Zotter, Analysis and Synthesis of Sound-
ware Package for Directional Audio Data,” in Proceedings Radiation With Spherical Arrays, Ph.D. thesis, Uni-
of the Fortschritte der Akustik, pp. 1059–1060 (Berlin, Ger- versity of Music and Performing Arts, Graz, Austria
many) (2010 Mar.). (2009 Nov.).

178 J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar.

ENGINEERING REPORTS DIRECTIVITY OF MUSICAL INSTRUMENTS

[27] M. Pollow, Directivity Patterns for Room Acousti- [31] D. Schröder and M. Vorlaender, “RAVEN: A Real-
cal Measurements and Simulations, Ph.D. thesis, RWTH Time Framework for the Auralization of Interactive Virtual
Aachen University, Aachen, Germany (2014 Sep.). Environments,” in Proceedings of the Forum Acusticum,
[28] V. Pulkki, “Virtual Sound Source Positioning Using pp.1541–1546 (Aalborg, Denmark) (2011 Jun./Jul.).
Vector Base Amplitude Panning,” J. Audio Eng. Soc., vol. [32] M. Lemke, A. Hoelter, D. Ackermann, and S.
45, no. 6, pp. 456–466 (1997 Jun.). Weinzierl, “Physics-Informed Interpolation of Directional
[29] F. Brinkmann and S. Weinzierl, “AKtools — An Characteristics,” presented at the Forum Acusticum (Torino,
Open Software Toolbox for Signal Acquisition, Process- Italy) (2023 Sep.).
ing, and Inspection in Acoustics,” presented at the 142nd [33] D. Ackermann, C. Böhm, F. Brinkmann, and S.
Convention of the Audio Engineering Society (2017 May), Weinzierl, “The Acoustical Effect of Musicians’ Move-
eBrief 309. ments During Musical Performances,” Acta Acust. united
[30] AES, “AES Standard on Acoustics - Sound Source Acust., vol. 105, no. 2, pp. 356–367 (2019 Mar./Apr.).
Modeling - Loudspeaker Polar Radiation Measurements,” https://doi.org/10.3813/AAA.919319.
AES Standard 56-2008 (2019 Aug.).

THE AUTHORS

David Ackermann Fabian Brinkmann Stefan Weinzierl

David Ackermann received his M.Sc. degree in Audiocom- Fabian is a member of the AES, German Acoustical Soci-
munication and Technology from Technische Universität ety (DEGA), and European Acoustics Association (EAA)
Berlin, Germany, in 2015. Since then he has been a research technical committee for psychological and physiological
associate in the Audio Communication Group at the Tech- acoustics.
nische Universität Berlin, where his research is focused •
in the field of virtual and musical acoustics. His research Stefan Weinzierl is head of the Audio Communication
includes the investigation of the time-dependent behavior Group at the Technische Universität Berlin. His research
of natural acoustic sound sources and their auralization in is focused on audio technology, virtual acoustics, room
virtual acoustic environments. acoustics, and musical acoustics. With a diploma in physics
• and sound engineering, he received his Ph.D. in musical
Fabian Brinkmann received an M.A. degree in Com- acoustics. He is coordinating a master program in Audio
munication Sciences and Technical Acoustics in 2011 and Communication and Technology at Technische Universität
Dr. rer. nat. degree in 2019 from the Technische Univer- Berlin and has coordinated international research consor-
sität Berlin, Germany. He focuses on the fields of signal tia in the field of virtual acoustics (SEACEN) and music
processing and evaluation approaches for spatial audio. information retrieval (ABC DJ).

J. Audio Eng. Soc., Vol. 72, No. 3, 2024 Mar. 179