You are on page 1of 4

Predicting sound quality of violin

from its frequency response


P Baiant, J s ' t ~ ~ d n e&kA Melka
Abstract
Two procedures for predicting the overall sound quality of a violin on a one-dimensional scale from its frequency
response were developed. The frequency responses of 43 violins with highly different sound qualities (from
Strads to factory-made violins) were measwed. The used computer-aided measurement method was developed in
WZORT in 1987 and is based on the ideas described by Diinnwald in 1984. The same instruments were
subjectively evaluated by competent experts and classified into 5 classes of sound quality. Analysis of the
frequency responses made with respect to the results of subjective classification revealed, that the 1100 measured
values of the frequency responses could be reduced to only 18 band-level values without substantial loss of
information on the sound quality. The positions and widths of the found frequency bands are very similar to the
distribution of critical bands. The first prediction procedure was based upon Factor Analysis , and the second one
upon Stepwise Multiple Regression. While both procedures give slightly different results, they have satisfactory
predictive power for the estimation of overall violin sound quality on a one-dimensional scale.

LEARNING APPUCATI ON VERIFICATION

PREDICTION OF SUBJ. PREDICTION ERROR


EXTRACTION OF EATRIX I A I
RATING OF INSTR. N + l

I ~ L . .N I.' I A I ~ I x I . . .NI+IE~ yk+l -1 ~ 1 xN+II


~ 1 err = YN+I -YN+~

Figure I: The predicting violin sound quality involves three phases:


a) In the learning phase the data from acoustic measurement and subjective ratings of sound
quality are employedfor construction of a sound quality predictor.
b) In the application phase data jkom acoustic measurement of an unknown instrument are
usedfor predicting its sound quality.
c) In the verification phase subjective rating of a new instrument is compared with its
predicted sound quality.
Predicting sound quality of a musical instrument The instruments were subjectively rated into 5
from its objective measured acoustic data is one classes, from excellent (class 1) to very bad (class
of the most general problems of music acoustics. 5) by competent violin makers and players.
Having sufficient number of relevant objective Numbers of instruments in class 1 to 5 were 8,
descriptions of the violin sound and corresponding 13, 8, 8, 6, (Table I).
subjective judgements, we may simulate the Sound of artificially driven violins was picked
process of subjective evaluation of sound quality. up in a way similar to method proposed by H.
That is what we call phase of learning. Provided Diinnwald 1984, by one microphone in anechoic
we have a knowledge of the relevant acoustical room in distance 2.8 meters. Recently we use very
parameters of an unknown instrument we are able efficient electrodynamic driver developed in
to predict the judgement of an average listener as Institute f i r Wiener Klangstil in Wienna. About
well as the judgement of a specific group of 1000 frequency response levels measured by a
listeners or of one listener with individual computer controlled system in the range 180 Hz to
perception model of instrument sound quality. 8000 Hz were transformed into 1/12 octave
The method was applied on a set of 43 violins, bands. The necessity of narrow band analysis
containing instruments from Strad to factory made demonstrate Fig. 2 (frequency response of a violin
ones. in octave, Bark, 113, 1/12 octave bands and in full

Table I:
List of 43 violins in experiment set, subjective classzJication, results of SMR, FA and
combined SMR+FA prediction of sound quality of unknown violin and errors of predicted
classification
violin class SMR FA SMR+FA e(SMR) e(FA) e(SMR+FA)
F090.Strad 1 .9 1.4 1 0 0 0
F09 1.Ruggie.r 1 1.3 1.1 1 0 0 0
F095.Schonbac 4 3.7 4.0 4 0 0 0
F096.Prokop 4 4.0 2.9 4 0 1 0
F097.Eberle 2 1.6 1.7 2 0 0 0
F098.Steiner 2 2.0 2.1 2 0 0 0
F099.Kuncer 3 1.9 2.3 2 1 1 1
F100.xl1.1 3 2.5 2.6 3 0 0 0
F101.bile.28 5 5.2 5.2 5 0 0 0
F102.bile.24 5 5.2 5.0 5 0 0 0
F103.Gala.47 4 3.1 3.8 4 1 0 0
F105.xO9 4 4.0 4.3 4 0 0 0
F113.Klotz 2 1.9 1.7 2 0 0 0
F114.Salteri 2 1.8 2.6 2 0 1 0
F115.Muschl 3 3.4 2.9 3 0 0 0
F117.Dvor.st. 4 3.9 3.6 4 0 0 0
FI 19.Guar.Sn 2 2.9 3.0 3 1 1 1
F120.xl0.405 1 .9 1.5 1 0 1 0
F121.xl0.404 2 2.7 2.9 3 1 1 1
F122.xl0.398 2 3.5 2.4 3 2 0 1
F123.xl0.399 2 2.0 2.1 2 0 0 0
F125.Guannes 2 2.2 2.4 3 0 0 1
F126.bile.25 5 4.4 4.7 5 1 0 0
F127.bile.05 5 4.4 4.8 5 1 0 0
F128.bile.23 5 4.5 5.2 5 0 0 0
F129.Placht 3 3.4 2.9 3 0 0 0
F130.Cermak 3 2.5 2.2 3 0 1 0
F13 1.Heinicke 3 2.9 2.7 3 0 0 0
Fl32.Metelka 3 3.6 2.9 4 1 0 1
F133.J.Gagli 1 1.6 .7 1 1 0 0
F136.GuarPet 1 1.9 2.2 2 1 1 1
F137.F.Rugg 1 1.3 1.9 2 0 1 1
F138.Ciochi 3 2.1 2.2 2 1 1 1
F139.J.Guarn 1 1.4 1.6 2 0 1 1
F140.Bemad. 2 1.5 1.7 2 0 0 0
F141.bile.29 5 4.9 4.8 5 0 0 0
F142.chemie2 4 4.2 4.8 5 0 1 1
F143.chemie3 4 3.4 3.8 4 1 0 0
F144.chemiel 4 2.5 2.4 3 1 2 1
F146.Amadeus 2 2.4 3.0 3 0 1 1
F15O.Iva.. 2 2.2 1.9 2 0 0 0
F152.Irena 2 2.0 1.6 2 0 0 0
F154.Clara 1 .6 .7 1 0 0 0
Total number of errors 14 IS 13
resolution) and Fig. 3 (correlation of levels of
average violin response with subjective sound
quality in the same frequency bands as in Fig. 2).
To solve the equation X*A=Y describing the
relationship between acoustic data and subjective
judgements we used firstly the Stepwise Multiple
Regression method. E.g., solution using only five
representative bands give rather high coefficient
I 5 10
" Bark
28
of multiple correlation r=0.852. The predicted
values of subjective ratings are shown in Table I,
column SMR.
Similar results gave application of Factor
Analyses (FA method of Principal Components
and Varimax rotation) on the same data.
In FA model is large number of variables
I 38 1/12 octave @
described by small number of factors. Each factor
represents group of highly correlating variables
(1112 octave bands). Predicting function was then
defined as a sum of object coordinates in factor
space weighted by correlation of factors with
quality.
FA computes all intercorrelations between
.I9 .?S .5 I 2 1 4 5 6 band levels, creates the factor space and estimates
Reqwncy response Je [k*l
the factor scores. The instruments represent cases.
Figure 2: Only narrow band description The factor scores were computed as sum of factor
represents well important peaks of frequency loadings weighted with levels in frequency bands
response. of individual violin. The predicting function was
then constructed as a sum of factor scores
weighted with corresponding correlation with
subjective quality ratings. E.g., solution for 5
factors gave correlation of predicted values with
subjectivejudgements r=0.827.
To verify reliability and precision of predicted
values, we repeated the learning phase 43 times,
r 1.1 &I-
81.
8 . l
01-
I - - I I I m
I
each time excluding one subjective rating of
instrument, which.had to be predicted. Subjective
I 5 I0
rating value was then compared with the predicted
l5 Bark value. The results are in Table I, columns SMR
and FA. The mean of both values gives a bit less
errors. It is worth to say that difference between
subjective rating and predicted value may be due
to erroneous value of median used to sum
subjectivejudgements.

Conclusions
1) Frequency response of a violin measured in the
described way contents crucial part of sound
quality information, which being extracted
predicts the perceived sound quality more reliable
than subjectivejudgements.
2) Radiated sound levels of a good instrument in 8
.I9 .B .5 I Z 3 4 5 6 fixed frequency bands (Table 11) should respect
lean response I [LC1 following rules:
Figure 3: Correlation between levels in 1/12
- first cavity resonance should be lower than 280
octave bands and subjective sound quality Hz,
follows well the excursions of mean
- the valley above 300 Hz shouldn't be very deep,
frequency response of 43 measured violins.
- otherwise should be levels in bands IlEl quite
Table 11:
Frequency bands where levels correlate positively with sound quality (inside these bands
should be high level without large peaks and gaps), frequency bands where levels correlate
negatively with sound quality (inside should be levels as small as possible without large
peaks) and correspondingfrequency bands of voice formants.

Band Fixed frequency bands 1/12 octave Formant Gap


No. + No.
1. 190 - 600 1-21 IlEl
2. 600 - 700 22-23 G1
3. 700 - 1200 24-32 A
4. 1200 - 1800 33-39 G2
5. 1800 - 3000 40-48 E2
6. 3000 - 3300 49 G3
7. 3300 - 4000 50-53 I2
8. 4000 - 54- G4

stable as well as the levels in bands A, E2 and 12,


- levels in gaps G1, G2 and G3 should be as small
as possible;
3) Established fixed frequency bands IlEl to G3
apparently conform to speech formant regions.
4) We believe the proposed method is ready to be
used in studies of any other musical instrument,
even for prediction of overall quality, including
aesthetic and' mechanic criteria.

References
Dilnnwald, H. (1984): "Die Qualitiitsbestimmung von
Violinen mit Hilfe eines objektiven Verfahrens",
Dissertation, TU Aachen, 14-21.

You might also like