You are on page 1of 10

Applied Acoustics 70 (2009) 636645

Contents lists available at ScienceDirect

Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust

A sound quality-based investigation of the HVAC system noise


of an automobile model
Ricardo Penna Leite *, Stephan Paul, Samir N.Y. Gerges
Federal University of Santa Catarina, Department of Mechanical Engineering, Laboratory of Vibrations and Acoustics, Post Box 476, Florianpolis, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: Using methods and techniques of sound quality engineering, the noise of the heating, ventilation and air-
Received 8 January 2007 conditioning system (HVAC) of an automobile model was studied. Such noise has a great inuence on
Received in revised form 17 June 2008 vehicle acoustical comfort and on overall quality perception of a vehicle. The study was divided into
Accepted 18 June 2008
two steps. The rst step aimed to identify the most signicant attributes that contribute to the perception
Available online 6 August 2008
of similarity or dissimilarity of this kind of noise, using the paired comparison technique and correlation
of the results with psychoacoustic models. Loudness, spectral composition and tonality, represented by
PACS:
the psychoacoustic models of loudness, sharpness, tone-to-noise ratio and prominence were found to
43.50.Qp
43.66.Lj
be the most important dimensions for the perception of similarity and dissimilarity of HVAC-noise.
In the second step of the study a model to predict subjective response to HVAC sounds using the
Keywords: semantic differential technique was developed. In particular the perception of annoyance was studied
HVAC system and it is shown that the annoyance caused by the HVAC noise can be satisfactorily described by Zwickers
Sound quality stationary loudness model, provided that the HVAC noises do not present tonal components. The loud-
Similarity ness model also predicts scores on a quiet/loud scale. Both results conrm the power of the loudness
Paired comparison dimension and its model introduced by Zwicker for the overall quality of stationary broadband sounds
Semantic differential
without slow uctuations or tonal components. From the annoyance model developed in this study a
maximum acceptable loudness level for HVAC-systems can be determined.
2008 Elsevier Ltd. All rights reserved.

1. Introduction high ventilation speeds, especially when no air-conditioning unit,


which is very usual in Brazil, is installed in the vehicle. Under these
In recent years the contribution pattern of noise sources to the conditions, the contribution of the HVAC noise to the acoustical
overall interior noise in a vehicle has changed. A large number of comfort of the passenger is very signicant.
secondary sound sources, previously masked, have become percep- In addition to the inuence that HVAC noise has on the comfort
tible to passengers with the reduction principally of engine noise and quality perception it can modify the thermal comfort inside the
levels and must be considered by the acoustical engineer. These car due to the interaction of thermal and acoustical perception [2].
noises, nowadays, play an important role in the passenger comfort Although it is considered that the lower the level of the HVAC
and have a considerable inuence on the quality perception not noise, the greater the interior comfort of the car will be, research
only of the source but of the whole automobile. Some of these activities conrming, denying or detailing this premise are not eas-
sources are active only occasionally, like the electric seat adjust- ily found in the literature. It is not known perfectly how HVAC sys-
ment or electric power windows. Others, like the HVAC system, tem noise should behave, regarding consumer expectation;
are likely to be operating as long as the automobile is in use. This interaction between the noise and thermal perception; the numer-
system can operate with alterations to the air temperature and ous working conditions of the HVAC system; interaction between
velocity if it is in manual mode, or can change these parameters the engine and HVAC systems; masking of the HVAC noise by the
over a larger time window if it is in automatic mode, where these engine noise and vice-versa. Thus, a complete study should aim
parameters and the state of the compressor (turned on or off) are to understand and model consumer perception, so that a HVAC
adjusted according to the thermal situation inside the vehicle. All system can be appropriately designed.
these working conditions introduce a large variability in the HVAC This study aims to investigate the perception of the HVAC noise
noise. In hot climate conditions, as found in Brazil, it operates at of an automobile model using methods and techniques of sound
quality engineering [1]. Specic objectives are (1) to identify the
* Corresponding author. Tel.: +55 48 3331 9227.
most important psychoacoustic models to characterize HVAC noise
E-mail addresses: ricardopennaleite@gmail.com (R.P. Leite), st.paul@emc.ufsc.br and (2) to propose models to predict subjective response to HVAC
(S. Paul), samir@emc.ufsc.br (S.N.Y. Gerges). noise, particularly the perception of annoyance.

0003-682X/$ - see front matter 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apacoust.2008.06.010
R.P. Leite et al. / Applied Acoustics 70 (2009) 636645 637

This part of the study was limited to the investigation of the the same model, taken from the production line. A headphone/
ventilation system alone, that is, the air-conditioning and engine microphone headset (NoiseBook from HEAD acoustics), placed
were switched off during recordings. This was done for two rea- on a persons head at the drivers position, was used with inde-
sons, (1) most cars of this model are still sold without air-condi- pendent of direction equalization [3,4].
tioning, and (2) for reasons of simplication, otherwise a very Windows and doors were always closed; the driver seat posi-
large number of combinations need to be considered, i.e., different tion was the same in all recordings; all valves of the air outlets
thermal conditions (air-conditioner loads), different (idle) engine were opened, and the direction of air ow was set to the center di-
sounds, sounds from the whole car at different velocities on differ- rectly towards the oor Fig. 1 to prevent noise induced by air ow
ent surfaces, etc. Therefore, the sounds being studied here are gen- at the microphones. All combinations of fan speed and direction of
erated only by the electric motor and the air ow. Further air ow were recorded. As the HVAC system has four fan speeds
investigations should be carried out considering more complex and four directions of air ow, sixteen conditions were recorded
working conditions of the HVAC system. for each car. For later analysis all recordings were edited, resulting
The car model chosen to make the recordings was a newly in sound les of 5 seconds duration. Only this length was consid-
introduced compact car model, developed in Brazil and destined ered, after consulting the literature (see [5]). If the sound is too
for the Brazilian and European markets. The model was chosen long, the subject will be bored, if it is too short, the subject will
for several reasons, the three most important being: (1) the high not perceived the sound characteristics.
production volume, resulting in a large number of cars available
for recording at the end of the production line without interfering 2.2. Subjective evaluation
too much in the activities of the line, (2) the fact that this model is
also being exported to Europe, a market with high quality de- The aim of the subjective evaluation of similarity was to reduce
mands, and (3) the fact that the percentage of cars of this model the number of sounds presented to subjects and evaluated by them
sold without air-conditioning unit is signicant. in further procedures and to determine the perceptual dimensions
The study was divided into two steps. In the rst (see Section 2), that are important for similarity ratings of HVAC sounds. Using the
the HVAC noise of fteen cars of the same model in sixteen paired comparison technique and a ratio-scale (Stone or line-scale)
different arrangements of fan speed and direction of air ow was with the anchors Very similar and Not very similar (the Portu-
recorded. Consequently, a large quantity of sound les was guese expressions were Muito parecidos and Pouco parecidos), as
obtained (15 cars  16 arrangements = 240 sound les). To reduce shown in Fig. 2, similarity of sounds were scaled by subjects, mark-
the number of sounds to be characterized and analyzed, a subjec- ing on the line-scale the point that best represented their impres-
tive evaluation of similarity, using the paired comparison tech- sion of similarity.
nique, was carried out. By means of statistical analysis (scatter Prior to the subjective evaluation procedure the sounds were
plots and multiple linear regressions) of objective and subjective grouped into 5 clusters (Table 1). Clusters 14 were sets of sounds
data, and additional analysis of qualitative data, the most signi- at the same fan speed and direction of air ow, cluster 5 was a set
cant characteristics of this kind of noise were identied. of sounds of the same car, at different fan speeds and directions of
In the second step (see Section 3), new recordings were taken air ow. Clustering aimed to reduce the number of possible combi-
in twelve cars. Based on the underlying dimensions of similarity nations for the paired comparison technique and, accordingly, to
ratings of HVAC sounds a set of sounds was chosen to be char- reduce the time required for the subjective evaluation without lim-
acterized and evaluated using the semantic differential tech- iting the variability unacceptably. Clusters were chosen in such a
nique. Using multiple linear regressions relations between way they would embrace a high and a low fan speed, two HVAC
subjective results and psychoacoustic models were determined models (Simple/Heating was not included because only two cars
and models to predict subjective response to HVAC sounds are of this model were recorded), all directions of air ow, and there
proposed. It is shown that the annoyance caused by the HVAC would be similar and different sounds within each cluster, as there
noise can be satisfactorily described by Zwickers stationary is a high variability in the noise of cars of the same model even at
loudness model. the end of the production line [6]. Furthermore, the sounds with
The HVAC system of the car model under investigation has four higher tonality had to be included. The number of sounds in each
fan speeds and four air outlets, here called Front, Windshield, Foot cluster depended on the number of cars of a model which were
and Foot/Windshield. Furthermore, three different kinds are available to be recorded.
available:

 Air-conditioning: with air-conditioning and heating


 Simple/Heating: without air-conditioning, but with heating
 Simple: without air-conditioning and heating

In this paper, the sounds will be referred to as, for example, car
002, Front-4. This means the recording of the HVAC noise of the car
002, working at fan speed 4 and air outlet Front.

2. Subjective evaluation of similarity

This section shows how the most important psychoacoustic


models to characterize the perception of the vehicle HVAC noise
were determined.

2.1. Recording and editing

Binaural recordings of ventilation noise (air-conditioning and Fig. 1. All valves of the air outlets were opened, and the direction of air ow was set
engine were always switched off) were made in fteen cars of downwards.
638 R.P. Leite et al. / Applied Acoustics 70 (2009) 636645

togram for pair 9. Of the subjects, 44% marked between zero and 2,
while 22% of subjects marked between 8 and 10. However, the
score distribution of some pairs looks like the normal distribution
Fig. 2. The ratio-scale used in the rst subjective evaluation.
or half of it. Thus, nonparametric statistical methods t best the
data of this study.
Analysis of the mean scores of the repeated pairs showed that
Table 1 subject responses were coherent. The equal pairs 3 and 34 had
The ve clusters of the evaluation mean values of 7.2 and 6.7 and standard deviations of 2.2 and
Cluster Model Working conditions Number of 2.4, respectively. In the same way, the equal pairs 7 and 28 had
sounds mean values of 7.2 and 7.5 and standard deviations of 2.4 and
1 Simple Foot-4 5 2.7, respectively. Application of the nonparametric Friedmans test
2 Air-conditioning Foot/Windshield-4 8 revealed that there were no signicant differences between the
3 Air-conditioning Front-2 8 means of the repeated pairs. Nevertheless, individual responses
4 Simple Windshield-2 5
5 Simple Various conditions of the same car 7
of many subjects diverged considerably between one pair and its
repetition.

2.4. Psychoacoustic properties responsible for perception of similarity


of HVAC noise
One sound in each cluster was chosen to be the reference and all
of the others were compared only with the reference, that is, the In this section, correlations between the psychoacoustic models
pairs were made up of the reference, which was played rst, and and the perception of similarity between sounds of the HVAC sys-
one of the other sounds of the cluster. Moreover, the references tem are investigated.
of clusters 1 and 2 formed a pair and this was also the case for
the references of clusters 3 and 4. 2.4.1. Analysis of subjects comments
To check the coherence of the subjects responses, two sound The questionnaire had a space for comments. Subjects usually
pairs were repeated. To verify if the presentation order is impor- wrote about loudness (in Portuguese, volume), level (nvel); or
tant, the pairs formed by references of clusters 1 and 2 and the ref- they wrote that one sound was duller (abafado, grave) or sharper
erences of clusters 3 and 4 were presented in the two possible (agudo) than the other; or they wrote about tonal components
orders. In total, the evaluation had thirty-four pairs. To familiarize (componentes tonais). Hence, it was concluded that loudness,
subjects with the test, a step called Training was included, which sharpness and tone-to-noise ratio are important psychoacoustic
used three additional sound pairs. models to characterize the similarity/dissimilarity of HVAC noise.

2.2.1. Subject proles and conversion of responses to numbers 2.4.2. Computation of the multiple linear regression variables
Twenty seven voluntary subjects took part in the evaluation. The well documented psychoacoustic models loudness, sharp-
The mean age was 29.7 years (standard deviation: 9.7 years) and ness, tone-to-noise ratio, prominence, roughness and uctuation
82% were male. Seventy-four percent of the subjects said they strength (see [7]) were applied to all sounds using dBSONIC soft-
owned a car or often drove their family car, 74% said they worked ware and diffuse eld parametrisation. Although subjects did not
with acoustics, and this means they were Masters or doctoral stu- make comments about the roughness or uctuation of the sounds,
dents in acoustics or undergraduted students that work on the pro- Brazilian Portuguese speaking subjects have difculty verbalizing
jects of the laboratory. On the one hand, this is a limitation, as the subjective impressions on sounds (see also [810]), these models
jury does not represent reliably all customers of this category of were included as they are very commonly employed in sound qual-
car, although all subjects are potential customers of this car. On ity investigations.
the other hand, it was expected that these subjects would be more As the aim of the subjective evaluation is to investigate the sim-
critical or at least have more ability to express their compliants ilarity perception, the difference modulus for the values of the
usually subjects face diculties in rating sound even when perceiv-
ing the sound characteristics, and this might be due to the limita-
tions of Brazilian Portuguese regarding the availability of words
and expressions to describe sound impressions, when compared
with languages like German. 12
Subjects were seated in the drivers position of a real car, which
was parked inside the garage of the laboratory. Sounds were pre- 10
sented to them via electrodynamic headphones.
Number of subjects

Subjective responses on the line-scale were converted to corre-


8
sponding numerical values between zero and ten, as the line had
10 cm. The value zero means that the subject marked the Not very
similar anchor and the value ten the Very similar anchor. Mean 6
values were calculated for each sound.
4
2.3. Response distributions and analysis of repeated pairs

2
To apply parametric statistical methods, assumptions about the
population scores and the sample randomness are required. These
assumptions are suitable for experimental sciences, but in many 0
0 2 4 6 8
cases not for behavioral ones. In this study, the response distribu-
Individual subject responses
tions showed that the population normality hypothesis can not be
applied. As an example, Fig. 3 shows the subjective responses his- Fig. 3. Histogram of Individual subject responses for pair 9.
R.P. Leite et al. / Applied Acoustics 70 (2009) 636645 639

psychoacoustic models for the sounds of each pair were calculated. 10


For example, if the loudness values of the paired sounds were 23

Subjective response (similarity)


and 26 soneGD, Dloudness j23  26j 3 soneGD. In this way,
the variables Dloudness, Dsharpness, Dfs (for uctuation strength), 8
Droughness, Dtnr (for tone-to-noise ratio), Dprominence and Dmpf
(the most prominent frequency) were created for each pair.
To investigate whether spectral composition inuences the sim- 6
ilarity perception, besides working with Dsharpness, the loudness
spectrum of each sound was divided into six bands, according to
4
Fig. 4. The mean loudness of each band was calculated. The differ-
ence modulus of the loudness bands for the sounds of each pair
was calculated, and the variables Dloud I , DloudII , DloudIII , DloudIV , 2
DloudV and Dloud VI were created.

2.4.3. A scatter plot investigation 0


Scatter plots are a way to investigate the correlation between 0 5 10 15 20 25
Loudness difference modulus (soneGD)
objective and subjective data. In Fig. 5, the loudness difference
modulus for the sounds of each pair is plotted on the abscissa Fig. 5. Relation between loudness difference modulus for the sounds of each pair
and subjective responses are plotted on the ordinate. According and subjective response.
to this gure, large differences in loudness for the sounds of a pair
caused a low score in the subjective evaluation, that is, subjects
perceived the sounds as being not very similar. However, there
are three pairs that received scores between 1 and 2 without hav- 10
ing a loudness difference greater than 5 soneGD. In the same way,

Subjective response (similarity)


there are pairs with a loudness difference of around 5 soneGD that
received greater scores, which means that loudness difference is 8
not the only important factor for perception of similarity.
This analysis was also carried out for roughness and promi-
nence. An interesting result is shown in Fig. 6: the more dissimilar 6
the perception of the sounds, the greater the sharpness difference
between the sounds of a pair.
4
Scatter plots show no correlation between subjective responses
and uctuation strength or tone-to-noise ratio.
2
2.4.4. Correlation of data with multiple linear regression (MLR)
The aim of multiple linear regression is to achieve equations
that link objective data (independent variables, values for psy- 0
choacoustic models) with subjective data (dependent variables, 0 0.1 0.2 0.3 0.4
Sharpness difference modulus (acum)
scores on similarity scale). The computation of the objective data
is explained in Section 2.4.2 and in Section 2.2 for the subjective Fig. 6. Relation between sharpness difference modulus for the sounds of each pair
data. Stepwise regression was applied and different equation mod- and subjective response.
els were tested. The rst model is shown in Eq. (1), where y is the
prediction of the subjective response. Note that, for identical
gree even if this pair had identical sounds.1 The scale with an
sounds, the equation will not yield ten, as would be expected as
unnumbered line reduces this problem (see [5]), but does not com-
this was the value attributed to the maximum degree of similarity.
pletely avoid it (see [11]). Unfortunately, this research did not use
In fact, the jury would not evaluate a pair with the maximum de-
pairs with identical sounds and thus it is not possible to verify a sub-
jects reaction to them. Eq. (1) is a model for the subjective re-
sponses: if the jury did not rate any pair with ten, the model will
represent it.
y 8:5  0:17  Dloudness  12:0  Dsharpness  0:58  Dtnr
 0:73  Dprominence  0:0003  Dmpf
1
2
The model gives R 0:84 using Dloudness, D sharpness, Dtnr,
Dprominence and Dmpf as independent variables.
In a second model (Eq. (2)), which gives R2 0:85, Dloudness
and Dsharpness are replaced by DloudVI . This is possible because
Dloudness and DloudVI are highly correlated with one another
R2 0:87 and Dsharpness and DloudVI are correlated R2 0:69
as well. This exchange cannot be carried out with variables

1
This is in agreement with the expected mean of the subjective data collected with
a limited scale. Some subjects may rate even identical sounds with positions that
correspond to values lower than the maximum degree. As they can not overstate the
Fig. 4. The division of loudness spectrum into six bands. maximum, the mean subjective rating is likely to be less than ten.
640 R.P. Leite et al. / Applied Acoustics 70 (2009) 636645

DloudI , DloudII , and DloudIII as they are not correlated with Dsharp- 10
ness (they are related to low and medium frequencies of the
spectrum).
8
y 8:3  9:4  DloudVI  0:45  Dtnr  0:97  Dprominence

Model response
 0:0002  Dmpf 6
2
Fig. 7 shows the subjective responses on the abscissa and the pre- 4
diction of these responses, according to Eq. (1), on the ordinate.
The straight line indicates the cases where model responses are
2
equal to subjective responses. Fig. 8 shows the same data, but
according to Eq. (2).
As the sounds of clusters 1 and 2 were for fan speed 4, their 0
loudness values were similar. Therefore, MLR was applied to the
pairs of clusters 1 and 2. A high correlation R2 0:91 is achieved 2
with only one variable (Dsharpness) as given in Eq. (3). Fig. 9 shows 0 2 4 6 8 10
the subjective responses on the abscissa and the prediction of Subjective response (similarity)
these responses, according to Eq. (3), on the ordinate.
Fig. 7. Comparison of subjective responses for similarity of all clusters and the
y 8:4  24  Dsharpness 3 prediction of these responses by Eq. (1). The straight line indicates the cases where
model responses are equal to subjective responses.
MLR was applied also to the pairs of clusters 3 and 4, as the sounds
of these clusters were all for fan speed 2 and their loudness values
were similar. However, contrary to the MLR results of clusters 1 and
2, Dsharpness is not correlated with subjective responses of clusters
3 and 4 R2 0:2. Instead, the model uses DloudIII and DloudV as 10
shown in Eq. (4). Fig. 10 shows the subjective responses on the ab-
scissa and the prediction of these responses, according to Eq. (4), on
8
the ordinate.

y 7:8  9:8  DloudIII  17  DloudV


Model response

4 6
MLR was also applied to the pairs of cluster 5. This cluster has four
sounds for fan speed 4 (including the reference) and one sound for 4
each fan speed 1, 2 and 3. Thus, the loudness values of these sounds,
estimated with Zwickers model, were very different. A high corre- 2
lation R2 0:95 was achieved only with Dloudness, as shown in
Eq. (5). Fig. 11 shows the subjective responses on the abscissa and
the prediction of these responses on the ordinate.
0

y 7:3  0:33  Dloudness 5


2
0 2 4 6 8 10
Subjective response (similarity)
2.4.5. Prediction of new observations
MLR can give equations that t the data almost perfectly, but Fig. 8. Comparison of subjective responses for similarity of all clusters and the
this does not mean that these equations will be appropriate to pre- prediction of these responses by Eq. (2).
dict new observations. In order to test whether Eqs. (1)(5) are able
to predict the results of new subjective evaluations of similarity,
the prediction interval [12] was determined. The pairs of sounds
that were used for Training were considered to be new observa-
tions, as this data did not form part of the MLR analysis. 8
For all pairs of the Training session, independent variables
7
(Dloudness, Dsharpness and so on) were calculated according to
the procedure of Section 2.4.2. The predictions of the subjective re- 6
sponses are given by Eqs. (1)(5) and they are presented together
Model response

with the minimum and the maximum values of the 95% prediction 5
interval in Table 2 for each Training pair. The table also shows the
subjective response obtained in the Training session. 4
From Table 2, it is clear that the predictions given by (1) and (2)
3
are closer to subjective responses than others and all subjective re-
sponses are within the prediction intervals. 2
Note that Eq. (4) can not be applied to pair Training 1 as it is de-
rived from the sounds of fan speed 2 and this pair had the sounds 1
of fan speed 4. Similar considerations were taken into account for
the pairs Training 2 and Training 3. 0
0 2 4 6 8 10

2.4.6. Partial conclusions


Subjective response (similarity)
The most important psychoacoustic properties to model the Fig. 9. Comparison of subjective responses for similarity of clusters 1 and 2 and the
perception of similarity of HVAC noise were found to be loudness, prediction of these responses by Eq. (3).
R.P. Leite et al. / Applied Acoustics 70 (2009) 636645 641

8 When the loudness difference between two HVAC sounds is suf-


ciently small, factors other than loudness become important for
7 similarity perception. MLR applied to the pairs of clusters 1 and
2 (all with sounds of fan speed 4 and similar loudness) yielded
6 Eq. (3), Dsharpness being the single independent variable. MLR ap-
Model response

plied to the pairs of clusters 3 and 4 (fan speed 2, similar loudness)


5 yielded Eq. (4), DloudIII and DloudV being the independent vari-
ables. These results suggest that, for small loudness differences,
4 the spectral composition is the most important factor. The impor-
tance of each frequency band changes according to the loudness le-
3 vel of the sounds being compared. When comparing sounds with
higher loudness, sharpness is a suitable psychoacoustic model for
2 estimation of similarity. But, for comparison of sounds with lower
loudness, a combination of two loudness bands is more suitable.
1 Most sounds under study did not have pure tone components,
0 2 4 6 8 10 thus it is believed that the inclusion of Dprominence and Dmpf in
Subjective response (similarity) some equations contributed to characterize timbre indicating
peaks in the spectrum and the frequency of a maximum peak.
Fig. 10. Comparison of subjective responses for similarity of clusters 3 and 4 and
Although the variable Dtnr does not have a linear relation with
the prediction of these responses by Eq. (4).
subjective responses, it is an appropriate parameter to characterize
the perception of similarity and is included in Eqs. (1) and (2). This
is in agreement with the fact that subjects wrote about tonal com-
ponents in their comments.
8 Fluctuation strength has no correlation with similarity percep-
tion. Roughness has a high correlation with loudness (between
Droughness and Dloudness, R2 0:85), so it is not necessary for it
6
to be included in models of perception of similarity.
Model response

4 3. A semantic differential approach

In the second step, it is shown that the annoyance caused by the


2 HVAC noise can be satisfactorily described by Zwickers stationary
loudness model.

0 3.1. Recording and editing

In order to collect samples from a different stage of the produc-


2
0 2 4 6 8 tion line and with a recording system that better models binaural
Subjective response (similarity) transfer functions at higher frequencies (this system was not avail-
able at the beginning of the project), new recordings were taken in
Fig. 11. Comparison of subjective responses for similarity of cluster 5 and the
prediction of these responses by Eq. (5). twelve cars of the same model, using a manikin (Mk 24-1 of 01 dB)
with diffuse eld equalization [3,13]. It was placed on the drivers
seat and a person was seated in the passengers seat to control the
sharpness, tone-to-noise ratio and prominence. When the loudness HVAC system, as shown in Fig. 12. To avoid background noise gen-
difference between two HVAC sounds is sufciently large, loudness erated by the hard-drive of the notebook, used as the data storage
was found to be the only important factor for perception of similar- device, the notebook was placed outside the car.
ity, as shown by Eq. (5), which comes from a cluster that had a Precautions taken to record the noises in the rst step, regard-
great loudness difference between its sounds. Furthermore, some ing adjustment of windows, doors, seat position and HVAC system,
subjects comments suggested that, in the case of a large difference were repeated. All combinations of fan speed and direction of air
in loudness, the rating of similarity will tend towards not very ow were recorded. Only the ventilation system was recorded, that
similar though subjects perceived the spectra of the sounds as is, the air-conditioning and engine were always switched off. For
similar. later analysis all recordings were edited. Post-editing sound les

Table 2
Prediction of the subjective responses of the pairs Training 1 to Training 3 and the minimum and maximum values of the 95% prediction interval

Equation Pair Training 1 Pair Training 2 Pair Training 3


Min Pred. Max Min Pred. Max Min Pred. Max
(1) 4.9 7.0 9.1 2.4 5.0 7.5 1.8 0.9 3.7
(2) 5.7 7.7 9.7 2.5 4.9 7.3 2.2 0.0 2.1
(3) 5.0 5.9 6.9
(4) 4.8 5.8 6.9
(5) 6.5 7.1 7.7 6.5 7.1 7.7 2.0 2.6 3.2
Subjective response 7.3 4.5 0.4
642 R.P. Leite et al. / Applied Acoustics 70 (2009) 636645

have a duration of 5 s, a fade in of 60 ms and fade out of 80 ms. Table 3


These values were chosen subjectively by the researchers. Bipolar adjective pairs and their respective category

Category Adjective pair Approximate translation


3.2. Development of the semantic differential Quality No-incmodo/ incmodo Not-annoying/annoying
Loudness Silencioso/barulhento Quiet/loud
The aim of the subjective evaluation was to determine the psy- Spectral composition Grave/agudo Dull/sharp
choacoustic models that can be used to describe the annoyance Roughness Suave/spero Smooth/rough
Tonality No-assobiante/assobiante Not-whistling/whistling
caused by the HVAC noise. To investigate the subjective impression
of HVAC noise, the semantic differential technique [14] was cho-
sen. Unfortunately, for Brazilian Portuguese no studies on auditory
descriptors or semantic differentials for sound evaluation are avail-
able [8], but this problem would have affected other commonly
used methods like paired comparison or magnitude estimation to
a similar degree. Thus, a simplied semantic differential design
was developed in various steps. The rst includes an open ques-
tionnaire applied to thirteen people regarding characteristics of
the noise of an automobile ventilation and air conditioning system
and their importance. The questionnaire was applied without and
Fig. 13. English translation of the semantic differential used.
with presentation of a selected set of four HVAC noises (two with a
higher loudness, but from different HVAC models, one a little qui-
eter and one with a low loudness).
Adjectives and nouns used to describe subjective impressions of tion [13]. Fluctuation strength was not calculated, because this
the presented sounds were collected in listening tests with naive model does not characterize well the kind of noise under study,
subjects and experts on acoustics and divided into four categories: as indicated by the rst step of this study (Section 2).
quality, loudness, roughness and spectral composition. For each Similarity of sounds was predicted using Eq. (1) and a total of 12
category, an adjective pair, as given in Table 3, was created. sounds that were predicted to be of low similarity were chosen.
Approximate translations are given in the second column. The pair Table 4 details the sounds used in the evaluation. To check
no-assobiante/assobiante (not-whistling/whistling) was added coherence of the subjects responses, two sounds were chosen
after the analysis of some German auditory descriptors used within to be repeated (Air-conditioning, Foot/Windshield-3 and Simple,
semantic differential designs of car interior sounds and aimed to Foot -4). In total, the evaluation had fourteen sounds. To familiarize
identify tonality. To make reading easier, Portuguese adjective subjects with the test, a step called Training was included once
pairs will be avoided in the text, approximate translations will be again, using two more sounds.
given instead. The rating scale used between the bipolar adjectives
had 7 degrees, as advised by Osgood, Suci and Tannenbaum [14]. 3.2.2. Subject proles and conversion of responses to numbers
Fig. 13 shows the semantic differential used in this study. Gradua- Thirty-one voluntary subjects took part in the evaluation. The
tion by adverbs was not used as no concept specic studies on mean age was 25.1 years, and standard deviation was 7.3 years.
semantic differentials for sound evaluation with Brazilian Portu- Ninety-seven percent were male and 61% said that they worked
guese have been carried out to date. In this case, it is recommended with acoustics. Eighty-seven percent said they owned a car or often
not to use graduation [15]. drove their family car. Subjects were seated in the drivers position
of a real car, which was parked inside the garage of the laboratory.
3.2.1. Criteria for sound selection Sounds were presented to them via electrodynamic headphones.
In order to chose sounds to be used in the listening test from the Subjects were not trained previously.
large number of sounds recorded (all HVAC system models, Subjective responses were converted to numbers between 3
fan speeds and air ow settings were recorded in all possible com- and 3, as shown in Fig. 14, and their means were calculated for
binations), criteria were established according to the results of the each sound.
previous similarity rating procedure. For all sounds, the psychoa-
coustic models loudness, sharpness, roughness, tone-to-noise ratio 3.3. Response distributions and analysis of repeated sounds
and prominence were calculated, using diffuse eld parametriza-
Once again, the response distributions showed that the
population normality hypothesis can not be applied. Hence, the
non-parametric Friedmans test was used to analyze the scores of

Table 4
The sounds used in the second subjective evaluation

Sound number Direction of air ow Fan speed HVAC model


1 Foot/Windshield 1 Air-conditioning
2 Windshield 2 Air-conditioning
3 Foot/Windshield 2 Simple/Heating
4 Foot/Windshield 2 Simple
5 Front 2 Air-conditioning
6 Windshield 3 Simple
7 Foot/Windshield 3 Air-conditioning
8 Windshield 3 Simple/Heating
9 Foot/Windshield 4 Simple/Heating
10 Foot 4 Simple
11 Front 4 Air-conditioning
12 Windshield 4 Simple/Heating
Fig. 12. Manikin position.
R.P. Leite et al. / Applied Acoustics 70 (2009) 636645 643

the sounds that were repeated to check subjects coherence. It was 5

Difference between notannoying/annoying


concluded that there were no signicant differences between the
means of the sounds that were presented twice. Again, individual
responses usually diverged between one sound and its repetition. 4

3.4. Comparison of results of the rst and the second subjective


evaluation 3

To analyze the coherence between the results of the rst and


second subjective evaluation, a scatter plot was created. The simi- 2
larity index for the sounds of the second evaluation, according to
Eq. 1, is plotted on the abscissa and the difference between re-
1
sponses on the not-annoying/annoying scale is plotted on the ordi-
nate. For example, the similarity index for sounds 1 and 2 (Table 4)
is 3.7. On the not-annoying/annoying scale, mean rating was 2.0 0
for sound 1 and 0.8 for sound 2. Thus, the difference of the means 10 5 0 5 10
is j2.0  (0.8)j = 1.2. These two values (3.7 and 1.2) were plot-
Similarity index
ted. The same procedure was carried out for all possible combina- Fig. 15. Comparison of results of the rst and the second subjective evaluation.
tions of pairs of the sounds of this evaluation, resulting in the
scatter plot shown in Fig. 15. The plot indicates that the more sim-
ilar the subjects perception of the sounds (higher values of the In the case of dull/sharp scale, the correlation with sharpness is
similarity index), the lower the difference between the responses weak R2 0:62. There might be two reasons for this: (1) To eval-
for these sounds on the not-annoying/annoying scale R2 0:73. uate this dimension subject training might be required, and (2) the
Portuguese descriptors might be ambiguous. For example, agudo
3.5. Correlations between subjective data (sharp) can also be understood as a musical note with a high pitch,
like an opera singer (in the words of one subject), and not as a
To investigate the relationship between the bipolar adjective broad-band noise where sharpness relates to the noise spectral
pairs of the semantic differential, their correlation was calculated. envelope characteristics.
It was found that the scales not-annoying/annoying, quiet/loud Because of its strong correlation with loudness and sharpness,
and smooth/rough are highly correlated with one another, as roughness can be excluded from models created to predict subjec-
shown in Table 5.
For the noises under study, the psychoacoustic proprieties loud-
ness and roughness were highly correlated R2 0:91. This may Table 5
explain the high correlation between the scores of the quiet/loud Matrix of R2 coefcients between subjective responses for the semantic differential
and smooth/rough scale. Probably due to this, subjects also re- pairs
ported having some difculties using scales without training. This Quiet/ Dull/ Smooth/ Not-whistling/
might have contributed to the similar scores on these scales. The loud sharp rough whistling
difculties reported by the subjects show the need for more basic Not-annoying/ 0.99 0.55 0.97 0.31
research on methods and techniques for sound quality assessment annoying
with Brazilian Portuguese speaking subjects. Quiet/loud 0.47 0.96 0.25
Dull/sharp 0.60 0.51
Smooth/rough 0.33
3.6. Correlations between psychoacoustic models

Table 6 is a matrix of R2 coefcients for correlations between Table 6


psychoacoustic models of the sounds under study. Sharpness, Matrix of R2 coefcients for correlations between psychoacoustic models of the
roughness and loudness are highly correlated. sounds under study

Sharpness Tone-to-noise ratio Prominence Roughness


3.7. Correlations between objective and subjective data
Loudness 0.85 0.09 0.28 0.91
2
Sharpness 0.09 0.27 0.88
Table 7 shows a matrix of R coefcients for correlations be- Tone-to-noise ratio 0.09 0.08
tween psychoacoustic models and data from subjective tests. It is Prominence 0.40
possible to see that not-annoying/annoying, quiet/loud and
smooth/rough scales are highly correlated with loudness, sharp-
ness and roughness. More than one model correlates with each
scale because of the high correlation between the psychoacoustic
Table 7
models (see Table 6). Matrix of R2 coefcients for correlations between psychoacoustic models and
The correlation presented in Table 7 becomes clearer in 16, a subjective results
scatter plot with loudness on the x-axis and the not-annoying/
Loudness Sharpness Tone-to- Prominence Roughness
annoying scale on the y-axis. noise ratio
Not-annoying/ 0.95 0.90 0.05 0.30 0.93
annoying
Quiet/loud 0.95 0.89 0.05 0.25 0.92
Dull/sharp 0.48 0.62 0.02 0.46 0.51
Smooth/rough 0.89 0.94 0.06 0.28 0.90
Not-whistling/ 0.16 0.26 0.04 0.21 0.25
whistling
Fig. 14. Seven step scale anchored at the bipolar adjectives.
644 R.P. Leite et al. / Applied Acoustics 70 (2009) 636645

3 3

2
2
notannoying / annoying

Model response
1
1
0
0
1

1
2

2 3
0 5 10 15 20 25 3 2 1 0 1 2 3
loudness (soneGD) Subjective response (notannoying / annoying)
Fig. 16. Scatter plot with loudness on the x-axis and the not-annoying/annoying Fig. 18. Comparison of subjective responses for the not-annoying/annoying scale
scale on the y-axis. and the prediction of these responses by Eq. (6). The straight line indicates the cases
where model responses are equal to subjective responses.

tive responses, and the scores on the smooth/rough scale can be


the quiet/loud scale is almost identical to Eq. (6), and for this reason
satisfactorily modelled using any of the remaining psychoacoustic
it is not shown here.
models.
Eq. (7) refers to the smooth/rough scale, which can be modeled
Fig. 17 presents a scatter plot with the tone-to-noise ratio on
using only roughness, with R2 0:90. The variable rou is the pre-
the x-axis and the not-whistling/whistling pair on the y-axis,
diction of the subjective response regarding the smooth/rough
showing no correlation between them. Subjects might have evalu-
scale.
ated this adjective pair reasonably if they had been trained. This
hypothesis is supported by an outlier at the top right of the scatter rou 5:0 0:23  roughness 7
plot: this sound has an audible tone and a higher value of the tone- Eq. (8) refers to the dull/sharp scale, which can be modeled using
to-noise ratio, and it was evaluated as the most whistling sound. only sharpness, with R2 0:62. The variable sha is the prediction
of the subjective response regarding the dull/sharp scale.
3.8. Multiple linear regression
sha 3:8 3:1  roughness 8
To obtain equations that relate subjective responses to psychoa- Note that the equations presented here are valid only for the kind of
coustic models, the multiple linear regression method was applied sounds under study: HVAC ventilation noise with low tonality.
to all adjective pairs, by means of stepwise regression. Eq. (6) refers
to the not-annoying/annoying scale, which can be modeled using
4. Conclusions
only loudness, with R2 0:95. The variable ann is the annoyance
provoked by HVAC noise. Fig. 18 shows the comparison between
In this study, methods and techniques of sound quality engi-
subjective responses for the not-annoying/annoying scale and the
neering were used to investigate the perception of the HVAC noise
prediction of these responses by Eq. (6).
of an automobile model. The main contributions of this work are
ann 2:5 0:24  loudness 6 (1) the identication of the most signicant attributes to character-
ize the perception of the HVAC noise, these being loudness, sharp-
As the not-annoying/annoying and quiet/loud scales are highly cor-
ness and the presence of tones (see Section 2.4.6); and (2) the
related with one another, an equation to predict the responses of
proposal of models to predict subjective response to HVAC noise,
the perception of annoyance, measured on a seven-point not-
1 annoying/annoying scale, being the most important. The derived
prediction model uses only Zwickers loudness and is shown by
Eq. (6). The loudness model can also be used to predict scores on
0.5 the quiet/loud scale. This result conrms the power of the loudness
notwhistling / whistling

dimension and its model introduced by Zwicker for the overall


0 quality of stationary broadband sounds without slow uctuations
or tonal components (the HVAC noise under study had low
0.5 tonality).
From Fig. 16 a maximum acceptable loudness level can be
determined. This information can be readily used in the automo-
1
tive industry and help engineers at car manufacturers and compo-
nent suppliers to select or improve HVAC parts to be installed in
1.5 the vehicle. It also conrms that the rst action to improve the
HVAC noise is to reduce its loudness level. However, these results
2 are limited to the ventilation system alone and apply only to a car
4 2 0 2 4 at rest with the engine switched off.
tonetonoise ratio (dB)
Future research should include more complex working condi-
Fig. 17. Scatter plot with tone-to-noise ratio on the x-axis and the not-whistling/ tions of the HVAC system and interaction with other sound
whistling pair on the y-axis. sources, such as engine noise. It should also include investigations
R.P. Leite et al. / Applied Acoustics 70 (2009) 636645 645

on other automobile models and HVAC noise characteristics, such References


as sharpness and tonality. For these investigations the loudness
of all sounds of a subjective evaluation should be equalized, other- [1] Leite RP. Estudo do rudo do sistema de ventilao automobilstica (study of
the noise of the automotive ventilation system). Masters thesis, Federal
wise the great importance of loudness in this kind of noise might University of Santa Catarina, Florianpolis, Brazil; 2006.
mask other psychoacoustic characteristics. Also, further studies [2] Roussarie V, Siekierski E, Viollon S, Segretain S, Bojago S. Whats so hot about
are necessary to introduce the required modications indicated sound? inuence of hvac sounds on thermal comfort. In: The 2005
international congress and exposition on noise control engineering
by this sound quality research. Another facet that should be inves- (internoise), Rio de Janeiro, Brazil, 710 August, 2005; 2005.
tigated is the inuence of the duration of the sound sample stimuli. [3] Fedtke Th, editor. Kunstkopftechnik eine Bestandsaufnahme. Report,
Even when recommended by the literature, or when there is a Normenausschuss Psychoakustische Messtechnik (NA 001-01-02-08 AK);
2007.
necessity to limit the duration of the assessment procedure, a short
[4] HEAD acoustics. Articial head equalization. Application note, 09/2006, in
duration of the sounds presented may inuence the subject rat- press.
ings, since no adaptation of their auditory system can occur. An [5] Otto N, Amman S, Eaton C, Lake S. Guidelines for jury evaluations of
investigation of this effect is therefore highly desirable and would automotive sounds. Sound Vib 2001;34(4):2447.
[6] Kompella MS, Bernhard RJ. Variation of structural-acoustic characteristics of
be of importance for the sound quality community. Major research automotive vehicles. Noise Control Eng J 1996;44(2):939.
should be carried out also in the eld of semantics for assessment [7] Zwicker E, Fastl H. Psychoacoustics. Facts and models. Berlin: Springer-Verlag;
of sound quality with Brazilian Portuguese speaking subjects. Even 1999.
[8] Paul S. A rst exploration of auditory descriptors for Brazilian Portuguese. In:
when terms like dull/sharp or not-whistling/whistling, that are The 2005 international congress and exposition on noise control engineering
commonly used with English or German speaking subjects, it is (internoise), Rio de Janeiro, Brazil, 710 August, 2005; 2005.
strongly recommended to revise the use of such terms in experi- [9] de Bitencourt RF, Paul S, de Andrade AL, Gerges SNY. Releva ^ncia dos
aspectos vibro-acsticos no conforto no interior de aeronaves. In: SOBRAC
ments with Brazilian speaking subjects, especially naive ones. (Ed.), Anais do XXI Encontro da Sociedade Brasileira de Acstica SOBRAC;
2006.
[10] Ribeiro YA, Leite RP, de Bitencourt RF, Lss Zmijevski TR, Gerges SNY. Uma
Acknowledgement abordagem do rudo de janelas eltricas atravs dos conceitos e tcnicas da
qualidade sonora. In: XXI Encontro da Sociedade Brasileira de Acstica e I
We would like to thank Volkswagen do Brasil LTDA, for helping Simpsio de Acstica de Salas, Edicaes e Escolas. Sociedade Brasileira de
Acstica SOBRAC; Novembro 2006.
this research to be carried out, the undergraduate students William
[11] Ilmoniemi M, Vlimki V, Huotilainen M. Subjective evaluation of musical
DAndrea Fonseca, Fernanda Higashi de Vasconcelos and Gustavo instrument timbre modications. In: Joint Baltic-Nordic acoustics meeting
Tabuti, for their help in carrying out recordings and subjective 2004, 810 June 2004, Mariehamn, land, Finland; 2004.
[12] Montgomery DC, Runger GC. Applied statistics and probability for engineers.
evaluations, and to all voluntary subjects. Further acknowledg-
3rd ed. United States of America: John Wiley & Sons, Inc.; 2003.
ments are given to Coordenao de Aperfeioamento de Pessoal [13] Daniel P, Equalization of articial head recordings. In: Proceedings of ICA,
de Nvel Superior (CAPES), to Conselho Nacional de Desenvolvi- Kyoto, Japan; 2004.
mento Cientco e Tecnolgico (CNPq) and to Financiadora de Estu- [14] Osgood CE, Suci GJ, Tannenbaum PH. The measurement of meaning. Urbana,
Chicago and London: The University of Illinois Press; 1975. 9th printing.
dos e Projetos (FINEP), for nancing the laboratory equipment and [15] Guski R. Psychological methods for evaluating sound quality and assessing
installations and also postgraduate scholarships. acoustic information. Acust Acta Acust 1997;83:76573.

You might also like