![Simple contrast of means across all perceptual attributes organized by condition and site. Contrasts by condition are between group at each site, whereas contrasts by site are within group for each condition. The scales for all attributes are normalised to the range [−1,1]. Significant differences as determined by posthoc contrast tests are accentuated](https://www.researchgate.net/publication/382080042/figure/fig2/AS:11431281259454254@1720496341180/Simple-contrast-of-means-across-all-perceptual-attributes-organized-by-condition-and_Q320.jpg)
Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Automating Urban Soundscape Enhancements with AI: In-situ
Assessment of Quality and Restorativeness in Traffic-Exposed
Residential Areas⋆
Bhan Lama,∗, PhD, Zhen-Ting Onga,Kenneth Ooia,Wen-Hui Onga,Trevor Wonga,
Karn N. Watcharasupatb,a, Vanessa Boeyc, Irene Leec, PhD, Joo Young Hongd, Ph.D.,
Jian Kange, Ph.D., Kar Fye Alvin Leef,g, PhD, Georgios Christopoulosf, PhD and
Woon-Seng Gana, PhD
aSchool of Electrical and Electronic Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore 639798, Singapore
bCenter for Music Technology, Georgia Institute of Technology, J. Allen Couch Building, 840 McMillan St NW, Atlanta, 30332, GA, USA
cBuilding & Research Institute, Housing & Development Board, Singapore 738973, Singapore
dDepartment of Architectural Engineering, Chungnam National University, 34134, Daejeon, Republic of Korea
eUCL Institute for Environmental Design and Engineering, The Bartlett, University College London, Central House,, 14 Upper Woburn
Place, London WC1H 0NN, United Kingdom
fNanyang Business School, Nanyang Technological University, 50 Nanyang Avenue, Singapore 639798, Singapore
gLaboratory of Neuropsychology and Human Neuroscience, Department of Psychology, The University of Hong Kong, Pokfulam Road, Hong Kong
ARTICLE INFO
Keywords:
urban soundscape
natural sounds
auditory masking
probabilistic approach
soundscape augmentation
artificial intelligence
ABSTRACT
Formalized in ISO 12913, the “soundscape” approach is a paradigmatic shift towards perception-
based urban sound management, aiming to alleviate the substantial socioeconomic costs of
noise pollution to advance the United Nations Sustainable Development Goals. Focusing on
traffic-exposed outdoor residential sites, we implemented an automatic masker selection system
(AMSS) utilizing natural sounds to mask (or augment) traffic soundscapes. We employed a
pre-trained AI model to automatically select the optimal masker and adjust its playback level,
adapting to changes over time in the ambient environment to maximize “Pleasantness”, a
perceptual dimension of soundscape quality in ISO 12913. Our validation study involving (𝑁=
68) residents revealed a significant 14.6% enhancement in “Pleasantness” after intervention,
correlating with increased restorativeness and positive affect. Perceptual enhancements at the
traffic-exposed site matched those at a quieter control site with 6dB(A) lower 𝐿A,eq and road
traffic noise dominance, affirming the efficacy of AMSS as a soundscape intervention, while
streamlining the labour-intensive assessment of “Pleasantness” with probabilistic AI prediction.
1. Introduction
1.1. Background and motivation
In urban environments, road traffic noise poses significant annual economic burdens, rivaling those of road
accidents, as evidenced by estimates in England (Âč7 billion) and across Europe (âĆň38 billion) [1,2,3]. Beyond
economic concerns, the documented adverse physical and mental health effects of urban noise warrant urgent mitigation
⋆The research protocols used in this research were approved by the institutional review board of Nanyang Technological University (NTU),
Singapore [IRB-2023-399].
∗Corresponding author
blam002@e.ntu.edu.sg (B. Lam)
ORCID(s): 0000-0001-5193-6560 (B. Lam); 0000-0002-1249-4760 (Z.-T. Ong); 0000-0001-5629-6275 (K. Ooi);
0000-0002-3878-5048 (Karn N. Watcharasupat); 0000-0002-0109-5975 (J.Y. Hong); 0000-0001-8995-5636 (J. Kang);
0000-0003-3774-6714 (K.F.A. Lee); 0000-0003-2492-653X (G. Christopoulos); 0000-0002-7143-1823 (W.-S. Gan)
Lam et al.: Preprint submitted to Elsevier Page 1 of 41
arXiv:2407.05744v1 [eess.AS] 8 Jul 2024
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
[1,4,5,6]. For instance, even a modest reduction of 5dB(A) in noise levels has been projected to yield substantial
annual economic benefits from adverse health effects in the United States, totaling $3.9 billion [3].
Crucially, mere reductions in sound pressure levels (SPLs) may not uniformly translate into perceptual improve-
ments. Considerable variations in annoyance and comfort levels have been found among individuals exposed to
identical SPLs, highlighting the complexity of the urban “soundscape” perception [7,8,9,10].
The soundscape approach, formalized in the ISO 12913 series [11,12,13], offers a holistic strategy for urban
sound management, aligning with the United Nations Sustainable Development Goals (SDGs), particularly SDG 3
(well-being) and SDG 11 (sustainable cities), by accounting for how humans perceive and experience their aural
environments, in context. The significance of this approach is echoed by the United Nations Environment Program
Frontiers 2022 report, which emphasized the need to mitigate unwanted noise while harnessing the health-promoting
benefits of natural sounds [14,15,16].
1.2. Soundscape augmentation for road traffic noise
Soundscape augmentation emerges as a viable intervention technique under the ISO 12913 paradigm. Additional
sounds, known as “maskers”, are augmented to existing soundscapes through loudspeakers or electroacoustic systems.
In prior art, maskers used in traffic-exposed urban areas typically comprise natural sounds, such as wind sounds [17],
sounds from animals (such as birds [18,19] and insects [20]), water sounds (such as man-made water features [21],
natural waterfalls [22], waves [23], and streams [24,25]), and corresponding mixtures [26].
Specifically, Calarco and Galbrun [27] modeled the propagation of water feature sounds in a park exposed to traffic
noise, defining optimal listening zones where water sounds were not less than 3dB below the traffic noise levels [22].
They found that the optimal zone decreases with increasing traffic noise levels, in addition to variations in preference
among various water feature varieties. Conversely, a laboratory study by Nilsson et al. [28] found a significant reduction
in traffic noise perception only when the fountain sound exceeded road noise by at least 10dB. A 9 % improvement
in overall sound quality post-augmentation was reported, favoring compositions with songbirds at varying volumes.
Furthermore, [29] found that participants were more likely to be highly annoyed when traffic noise was perceived
to be the dominant sound source under augmentation with birdsongs and stream sounds. On the contrary, a separate
virtual reality (VR)-study found no evidence that any particular birdsong composition augmented to soundscapes of
a Swedish park reduced stress levels [30]. Van Renterghem et al. [31] explored real-world soundscape augmentation
in a traffic-exposed park by inviting participants to customize natural sound samples emitted from a hidden speaker
to their preference. Hence, it would be naive to assume that every bird masker (or every masker from the same class
in general) would improve the quality of a given soundscape, thereby necessitating some form of selection process to
effect a desired perceptual change.
Lam et al.: Preprint submitted to Elsevier Page 2 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Moreover, few studies have extended their findings into soundscape augmentation systems for road traffic noise
in real-life urban environments. Installing and uninstalling speakers in a soundscape augmentation system can also
be more cost-effective and conducive to the surrounding environment as compared to alternative methods of noise
mitigation such as noise barriers, which require physical space and may be more difficult to retrofit to existing urban
areas [32].
1.3. Masker selection methods for soundscape augmentation
One real-life soundscape augmentation system was explored by Van Renterghem et al. [31] in a park in Ghent,
Belgium, where road traffic noise was dominant. Participants composed their own maskers by adjusting the playback
levels of eight natural sound samples emanating from a hidden loudspeaker, then evaluated both the original and
augmented soundscapes. The study observed a mean improvement of 0.36 unit (9%) in overall sound quality on a
5-point scale, with most participants preferring the sounds of house sparrows and mixed songbirds.
Similar effects may also be observed even if the loudspeaker or speaker systems are visible to the participants. Hong
et al. [33] conducted a study with participants standing at pedestrian walkways near roads, adjusting the soundscape-
to-masker (SMR) ratio of birdsong and fountain recordings reproduced by down-firing speakers of a mixed-reality
device. The recordings were accompanied either by a hologram matching their source (a bird for the birdsong and a
jet-and-basin fountain for the fountain) or by a visible speaker. Participants adjusted the SMR to a level they found
most preferable for masking traffic noise. The study found no significant differences in the chosen SMRs or the
resultant ratings of overall soundscape quality and perceived loudness of traffic noise between the hologram and speaker
conditions. In addition, Regazzi et al. [34] used the frequency spectrum of transformer noise in a residential area to
create a natural sound masker, aiming to equalize tonal frequencies when reproduced over speakers. This demonstrates
the effectiveness of speakers in soundscape augmentation, despite the potential lack of realism compared to real-life
sources.
However, these methods require participant involvement or expert input to generate optimal maskers and playback
gains, which may not be practical for long-term deployments. Changing soundscape characteristics over time can render
previously optimal maskers suboptimal.
Alternatively, model-based approaches offer the potential for generalizability across scenarios. For instance, Lenne
et al. [35] optimized masker playback locations indoors based on room acoustics simulations, while others have
incorporated physical models for real-time augmentation of footstep sounds in virtual-reality soundscapes [36,37,38].
Suhanek et al. [39] optimized the “total distraction coefficient” to select appropriate songs as maskers for park and
expressway soundscapes, but only theoretically validated their masker choices. Despite the promise, model-based
Lam et al.: Preprint submitted to Elsevier Page 3 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
approaches remain sparse in the literature, particularly in the context of road traffic noise, and none have been developed
using the ISO 12913 framework.
Automated masker selection methods could enhance efficiency by reducing the time and labor involved in human
evaluation, while also adapting to changing soundscapes. The success of automated masker selection relies on the
availability of reliable models to predict affective responses, such as “Pleasantness” (ISOPL) [13] or restorativeness
[40,41], which are crucial for enhancing acoustic comfort. To date, few prediction models for multidimensional
indicators such as ISOPL have been developed [42,43,44,45,46], and interventions based on enhancing ISOPL
are lacking [47].
1.4. Research questions
Addressing these gaps, we utilize our probabilistic ISOPL prediction model, trained on our large-scale dataset of
perceptual responses to soundscapes [48], to deploy and validate a proof-of-concept model-based automatic masker
selection system (AMSS) at a traffic-exposed residential site. Operating autonomously, the AMSS augments the
soundscape to maximize ISOPL. Through in-situ validation, we aim to assess the impact of AMSS on soundscape
quality, its influence on related perceptual dimensions, and its correlation with objective acoustic metrics. Specifically,
we seek to answer the following research questions:
RQ1. To what extent can the soundscape quality of a traffic-exposed site be modified by the AMSS?
RQ2. What impact does optimizing a soundscape intervention to improve ISOPL have on other soundscape-related
perceptual dimensions, such as restorativeness, perceived loudness, and ISOEV?
RQ3. How do perceptual changes induced by the AMSS correlate to objective (psycho)acoustic metrics?
2. Method
The in-situ validation study was conducted between 1 August 2023 and 30 November 2023, and prior to participant
recruitment and experimentation, formal ethical approval was obtained from the Institutional Review Board at Nanyang
Technological University (Reference number IRB 2023-399). The study administrators strictly adhered to the approved
methodology, and informed consent was obtained from all participants prior to the start of the experiment.
2.1. Study sites
The study sites were two distinct pavilions within a public residential estate in Singapore, as shown in Figure 1.
Both pavilions were identical in design, but were situated at different locations in the estate.
The first study site was a ground-floor (“GND”) pavilion positioned at street level adjacent to a children’s playground
and fitness area. The GND was situated amidst six residential apartment blocks, which were in turn surrounded by and
Lam et al.: Preprint submitted to Elsevier Page 4 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
served as a physical barrier to a minor 2-lane road (60 m away from the pavilion) with light traffic. As a control site,
no AMSS was deployed at the GND.
The second study site was a rooftop (“ROOF”) garden pavilion positioned near the periphery of rooftop garden atop
an 8-storey multi-storey car park (MSCP), which bordered a major 8-lane expressway with heavy traffic. The ROOF
was positioned 30m above street level and was flanked by a 2-lane slip road (50m away) leading out from a major
6-lane expressway (70 m away). The AMSS was physically deployed at the ROOF, with four loudspeakers (Moukey
M20-2, DONNER LLC, FL, USA) affixed to the pavilion roof (at a height of 2.5 m above the ground of the pavilion)
in a square of length 2.2 m for the playback of maskers, which were automatically selected and reproduced according
to the method described in Section 2.3. A customized Internet-of-Things (IoT)-based infrastructure was used for the
deployed AMSS, as detailed by [49]. The placement of the hardware of AMSS did not physically or visually block any
ingress or egress routes to the ROOF.
2.2. Design of in-situ validation experiment
To investigate the influence of the AMSS on soundscape perception, we employed a within-between design.
Participants were allocated randomly into two independent groups (between factor): the “AMSS” (AMSS) and the
“Ambient” (AMB) group. In both groups, participants evaluated the soundscapes at both the GND and ROOF (within
factor) in a randomized order. However, the AMSS was turned on (i.e., soundscape augmentation was performed
according to the method described in Section 2.3) for the AMSS group at the RTGP and turned off (i.e., no soundscape
augmentation was performed) for the AMB group at the ROOF. As explained in Section 2.1, the AMSS was not deployed
at the GND, so the evaluations at the GND for participants in both the AMSS and AMB groups corresponded to that of the
ambient environment at the GND. Given the communal nature of the public space, each session accommodated up to
four participants, aligning with the maximum seating capacity of the pavilions. On average, there were 1.53 ±0.80
participants per session, and an overview of the experimental procedure for each session is illustrated in Figure 2.
At the onset of each session, participants convened at the meeting point (MP) for the requisite consent process, a
briefing on the study protocol, and hands-on training with the electronic form used for the evaluations. The International
Positive and Negative Affect Schedule Short Form (I-PANAS-SF) [50] was also administered at the meeting point.
To prevent undue bias in evaluation, participants were not informed whether they had been placed into the AMSS or
AMB group, and were also not informed about the presence of the AMSS system at the ROOF. Within each study site,
participants were initially directed to listen to the pavilion’s soundscape for 10 min without engaging in any other
activities and without interacting with each other. Subsequently, they used their personal mobile devices to complete
an electronic evaluation form. To ensure clarity, study administrators reiterated the following instructions verbatim to
participants before the listening period:
Lam et al.: Preprint submitted to Elsevier Page 5 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
(a) (b)
Microphone captures
ambient sounds
Play maskers at
selected gain
Cloud subsystem
metadata {masker,gain}
Masker
log-mel
spectrograms
Randomized
log-gain
Pre-trained ISOPL
prediction model
masker_V
ma...
masker_U
masker_B
ISOPL
Masker with
highest ISOPL
score selected
Retrieve selected
maskerlocally
Log-mel
spectrograms
In-situ subsystem
(c)
Figure 1: Study sites in a public residential estate in Singapore: (a) A ground-floor pavilion (GND) in the outdoor recreational
area at coordinates (1.401358, 103.895427). (b) A rooftop garden pavilion (ROOF) situated atop an 8-storey multi-storey
car park at GPS coordinates (1.343373, 103.686134). (c) An overview of the end-to-end process of the automatic masker
selection system (AMSS)
We will be assessing the sound environment within the pavilion. Over the next 10 minutes, immerse yourself
in the surrounding sounds. Choose to sit or stand, but minimize movements to avoid disturbing others.
Refrain from using your phone or engaging in other activities. Focus on the types of sounds and your
emotional responses, considering the pavilion’s context for rest and relaxation.
During the 10 min listening period, the acoustic environment experienced by the participants was captured using a
binaural microphone (TYPE 4101-B, Hottinger BrÃijel & KjÃęr A/S, Virum, Denmark) equipped with a windscreen.
This microphone was coupled with a data recorder (SQobold, HEAD acoustics GmbH, Herzogenrath, Germany).
Lam et al.: Preprint submitted to Elsevier Page 6 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Figure 2: Overview of experimental procedure and data collected from participants for the in-situ validation experiment.
Ensuring data precision and uniformity, the binaural recording equipment underwent calibration using an IEC 60942
class 1 calibrator (42AG, G.R.A.S. Sound & Vibration A/S, Holte, Denmark). The responsibility of wearing and
operating this equipment rested with a single experiment administrator during each session, with a total of four unique
administrators overseeing the entire 4-month study duration.
In alignment with ISO 12913-2 [12], environmental data was systematically collected during the 10min listening
period. Temperature and humidity readings were obtained from a combined digital humidity and temperature sensor
(BME280, Bosch Sensortec GmbH, Reutlingen, Germany), while luminance data was captured by an optical sensor
(LTR-559ALS-01, LITE-ON Technology Corp., Taiwan) integrated into the AMSS system at the ROOF, all at 10min
intervals. Additionally, wind speed, 24-h pollutant standards index (PSI), and PM2.5 readings were sourced from
the nearest weather station via the Singapore Meteorological Service, also recorded at 10 min intervals. Detailed
specifications regarding the metrics, range, accuracy, and resolution of the measurement instruments are delineated
in Table 1.
After the 10 min listening period, the participants were instructed to complete the I-PANAS-SF questionnaire.
Thereafter, participants received the following instruction:
This evaluation is about the surrounding sound environment you just experienced in the past 10 minutes.
Answer the following questions by recalling the sounds you experienced in the 10 minutes.
The questions formed the site evaluation questionnaire, which prompted participants to rate (1) the dominance of
noise (DOMNoi), (2) the dominance of natural sounds (DOMNat), (3) the dominance of human sounds (DOMHum), (4)
the 8 attributes corresponding to perceived affective quality (PAQ) in the Method A questionnaire of ISO 12913-2, (5)
Lam et al.: Preprint submitted to Elsevier Page 7 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table 1
Critical specifications of measurement instruments.
Instrument Metric Range Accuracy Resolution/Sensitivity
TYPE 4101-B Binaural
Microphone
Sound pressure (Pa) 20 Hz âĂŞ
5 kHz
±𝟤 𝖽𝖡 re 1 kHz 20 mV/Pa ±𝟥 𝖽𝖡
5 âĂŞ 20 kHz 3 dB soft boost at 0°
incidence
BME280 Digital
humidity, pressure and
temperature sensor
Temperature (°C) 0 – 65 ±0.5 0.01
Relative humidity
(%RH)
0 – 100
(0 – 60°C)
±3 0.008
LTR-559ALS-01 Optical
Sensor
Luminance (lx) 1 64000 0.977
the overall soundscape quality (OSQ), (6) appropriateness (APPR), and (7) perceived loudness (PLN) on 5-point scales,
on top of the items in the 18-item Perceived Restorativeness Soundscape Scale by [41], which were on 7-point scales.
The PRSS consists of four main dimensions: Fascination (PRSSFas), Being-Away (PRSSBA), Compatibility (PRSSCom),
and Extent, which consists of two sub-dimensions: Extent-Coherence (PRSSEC) and Extent-Scope (PRSSES). The
precise wording of each item in the site evaluation questionnaire is provided in Appendix A, Table A.1.
Considering the fatigue and relevance of terms within the local context, the 18-item PRSS scale utilized in this
study underwent modification by omitting or consolidating 7 of the 23 items from the PRSS scale with specific framing
outlined by Payne and Guastavino [41]. These adjustments are detailed in Table A.2.
At the end of the soundscape evaluation at the second site, participants completed an additional participant
information questionnaire covering basic demographics (gender, age, occupation) and self-reported assessments on
the (1) individual noise sensitivity (INS) [51], (2) baseline noise annoyance (BNA) [52], (3) Perceived Stress Scale
(PSS-10) [53], and (4) WHO-Five Well-being Index (WHO-5) [54]. exact wording of every item in the participant
information questionnaire can be found in Table A.3. The experimental procedure averaged 53.41 ±11.81 min to
complete.
2.3. Stimuli and automatic masker selection
As explained in Section 2.2, only the AMSS group experienced augmented soundscapes with maskers presented
over four loudspeakers in the ROOF. The maskers were selected from the bank of maskers in the ARAUS dataset [48],
comprising 280 different processed recordings of birds, water, wind, traffic, and construction as 30s mono tracks.
Specifically, a pre-trained artificial intelligence (AI) model decoupling the spectrograms of the existing soundscape,
masker, and playback gain [44] was used in the AMSS to pick maskers and corresponding gain values, in intervals
of 30 s. The model was trained on the 25,440 subjective responses to augmented urban soundscapes in the ARAUS
dataset to predict distributions of ISO Pleasantness (ISOPL), as defined in ISO 12913-3 [13], which the AMSS then
Lam et al.: Preprint submitted to Elsevier Page 8 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table 2
Frequency distribution of the maskers chosen by the AMSS during the 10-min listening period across all “AMSS” group
participants. Description and availability of the corresponding maskers as detailed by Ooi et al. [48] in the ARAUS dataset.
Maskers Frequency (%) Description
bird_00012 0.2% Bahama Mockingbirda
bird_00025 1.0% Baltimore Orioleb
bird_00069 26% Northern Cardinalc
bird_00071 5.8% Veeryd
bird_00075 67% Common Redshanke
aPaul Driver, XC140239. Accessible at www.xeno-canto.org/140239.
bEric DeFonso, XC370500. Accessible at www.xeno-canto.org/370500.
cChristopher McPherson, XC601752. Accessible at www.xeno-canto.org/601752.
dChristopher McPherson, XC602571. Accessible at www.xeno-canto.org/602571.
eJoao Tomas, XC604437. Accessible at www.xeno-canto.org/604437.
used to select a masker-gain combination at each interval to maximise the ISOPL of the existing soundscape at the
ROOF. An overview of the AMSS system is depicted in Figure 1c.
At each 30s interval, the AMSS randomly picked 5 gain values from a log-normal distribution for each masker
candidate in the masker bank, with the log-gains being normally distributed with mean −2.0and standard deviation
1.5. These values match the distribution of log-gains in the ARAUS dataset maskers when calibrated to an SPL
of 65 dB(A) and correspond to five possible SMRs when applied to the maskers upon playback. For each of these
masker-gain combinations, the AI model gave as initial output the predicted ISOPL distributions as though they were
used to augment the existing soundscape. Then, the masker-gain combinations were ranked in terms of the predicted
improvement in ISOPL via the estimation scheme described by [45]. Lastly, the top-ranked masker-gain configuration
was reproduced across the four loudspeakers in the deployed AMSS with each loudspeaker playing back the same
masker at the same SPL corresponding to the gain value.
The sound level output of each loudspeaker was previously calibrated for each masker from 46 to 83dB(A)
in 3 dB(A) intervals using a custom automated procedure in a soundproof box [55], at a distance of 1 m from a
measurement microphone (146AE, G.R.A.S. Sound & Vibration A/S, Holte, Denmark). The desired output sound
level of the masker corresponding to the gain value determined by the AMSS was achieved by energetic interpolation
and compensation for distance (inverse square law) and number of speakers (4 speakers).
A total of 481 instances of maskers selected by the AMSS and reproduced over the loudspeakers at the ROOF
were logged across 18 of the 20 sessions during the AMSS condition. Maskers bird_00069 (26 %) and bird_00075
(67 %) were selected the most often, which were sometimes interjected by bird_00071 (5.8 %), bird_00025 (1.0 %)
and bird_00012 (0.2 %), as delineated in Table 2. The frequency of masker presentation in Table 2depicts an average
participant’s exposure during the 10 min listening period preceding the evaluation at the ROOF for the AMSS group.
Lam et al.: Preprint submitted to Elsevier Page 9 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table 3
Summary statistics of environmental parameters captured at ROOF during the 10-min listening period across all
participants.
Environmental Parameter AMSS1Ambient1p-value2
Temperature (°C) 31.64 (1.37) 33.39 (2.13) 0.083
Relative Humidity (%RH) 59.09 (4.20) 56.02 (6.96) 0.494
Luminance (lx) 314.65 (132.44) 334.45 (162.26) 0.750
Wind Speed (km h−1) 3.63 (0.76) 3.17 (1.25) 0.259
24-h PSI 50.17 (6.08) 45.50 (6.50) 0.203
PM2.5 (µg m−3) 16.00 (4.15) 11.92 (4.54) 0.098
1Reported as “Mean (Standard deviation)”
2Wilcoxon rank sum exact test; Wilcoxon rank sum test
2.4. Non-acoustic environmental conditions for in-situ validation study
As measured by the instruments shown in Table 1, the in-situ experimental conditions exhibited notable stability
across all parameters for both the AMSS and AMB groups, as presented in Table 3. The prevailing temperature and
humidity levels align with the characteristic hot and humid tropical climate of Singapore, complemented by wind
speeds indicative of light air. A noteworthy consideration is the absolute luminance levels, which were damped by the
tinting on the protective cover over the sensor.
Importantly, the air quality remained within healthy limits throughout the entire study duration. Employing
Wilcoxon rank-sum tests at a 5 % significance level revealed no significant distinctions between the AMSS and AMB
groups across key environmental parameters of temperature, relative humidity, luminance, wind speed, 24-hour PSI,
and PM2.5 readings. Given the inherent in-situ nature of this study, where environmental parameters are beyond
direct experimental control, the discovery of non-significant differences between groups is fortuitous but noteworthy.
This outcome allays concerns associated with potential confounding factors stemming from divergent environmental
conditions between the AMSS and AMB groups.
2.5. Participants
A cohort of 70 participants participated in this study. Recruitment was executed through mobile messaging
channels and the distribution of advertisements via grassroots organisations. The study inclusion criteria mandated
that participants reside within the designated postal sector of the study site (i.e., postal sector 82) and fall within the
age range of 21 to 70 years. Participants received remuneration in the form of supermarket vouchers with a value of
$30 (Singapore dollars).
Due to the onset of a thunderstorm midway through one of the study sessions, data from the two participants
for that session were deemed unreliable and subsequently excluded from the analysis. The final dataset comprised
responses from 68 participants, consisting of 40 females (59%) and 28 males (41 %), with a mean age of 41.75 and
Lam et al.: Preprint submitted to Elsevier Page 10 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table 4
Summary of participant demographics and non-acoustic factors (PSS-10, WNSS, WHO-5, baseline annoyance) across each
condition (AMSS and AMB).
Overall, N = 681Ambient, N = 321AMSS, N = 361p-value2
Gender 0.09
ÂăÂăÂăÂăFemale 40 (59%) 21 (66%) 19 (53%)
ÂăÂăÂăÂăMale 28 (41%) 11 (34%) 17 (47%)
Age 41.75 (12.83) 42.00 (13.22) 41.53 (12.65) 0.91
Occupation
ÂăÂăÂăÂăEmployed 48 (71%) 26 (72%) 22 (69%)
ÂăÂăÂăÂăOther 1 (1.5%) 1 (2.8%) 0 (0%)
ÂăÂăÂăÂăRather not say 4 (5.9%) 2 (5.6%) 2 (6.3%)
ÂăÂăÂăÂăRetired 6 (8.8%) 2 (5.6%) 4 (13%)
ÂăÂăÂăÂăStudent 6 (8.8%) 3 (8.3%) 3 (9.4%)
ÂăÂăÂăÂăUnemployed 3 (4.4%) 2 (5.6%) 1 (3.1%)
PSS-10 0.51 (0.13) 0.51 (0.13) 0.51 (0.14) 0.94
INS 0.67 (0.06) 0.67 (0.05) 0.67 (0.06) 0.72
WHO-5 0.62 (0.17) 0.59 (0.17) 0.65 (0.16) 0.54
BAaircraft 3.93 (1.39) 3.88 (1.41) 3.97 (1.38) 0.82
BAmrt 2.35 (1.22) 2.59 (1.29) 2.14 (1.13) 0.46
BAconsite 3.53 (1.30) 3.59 (1.29) 3.47 (1.32) 0.80
BAreno 3.46 (1.34) 3.59 (1.39) 3.33 (1.31) 0.59
BAtraffic 3.46 (1.20) 3.53 (1.14) 3.39 (1.27) 0.90
BAanimals 2.12 (1.10) 1.94 (1.05) 2.28 (1.14) 0.28
BAchildren 2.51 (1.17) 2.66 (1.21) 2.39 (1.13) 0.51
BApeople 2.34 (1.02) 2.47 (1.05) 2.22 (0.99) 0.28
BAothers 2.35 (1.18) 2.38 (1.10) 2.33 (1.26) 0.83
1Gender and occupation reported as “Count (%)”; all others reported as “Mean (Standard deviation)”; 2Four-sample test
for equality of proportions without continuity correction for gender, and Exact two-sample Kolmogorov-Smirnov test
otherwise
standard deviation of age of 12.83, as detailed in Table 4. Participants were generally working-class adults, but the
employment status varied among individual participants, with a majority being employed (71%), followed by retirees
(8.8 %), students (8.8 %), unemployed individuals (4.4%), and a segment that either did not disclose or fell into the
“other” category (7.4 %). The AMSS group comprised 36 participants, while the AMB group consisted of 32 participants.
Variations in proportions of gender between the AMSS and AMB groups were determined to be non-significant via four-
sample test for equality of proportions without continuity correction (𝑝= 0.09). On the other hand, the age was similarly
distributed between AMSS and AMB(𝑝= 0.91). The central tendencies and dispersion of the self-assessed PSS-10, INS,
WHO-5, and baseline annoyance across all noise categories were similar across both AMSS and AMB groups, as detailed
in Table 4.
Tests of distribution equality were performed across age, PSS-10, INS, WHO-5, and all baseline annoyance
categories, acknowledging the potential influence of non-acoustical factors on soundscape perception [56,57]. Analysis
using the exact two-sample Kolmogorov-Smirnov test revealed no significant differences between the AMSS and AMB
groups, as listed in the 𝑝-value column in Table 4.
Lam et al.: Preprint submitted to Elsevier Page 11 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
2.6. Data analysis
From the binaural recordings collected in Section 2.2, objective acoustic and psychoacoustic indices were computed
with a commercial software package (ArtemiS suite, HEAD acoustics GmbH, Herzogenrath, Germany) on the
representative channel with the highest value [13]. These included both the A- and C-weighted equivalent sound
pressure level over each 10min listening period (𝐿A,eq;𝐿C,eq), and the 95 % exceedance level of psychoacoustic
loudness (𝑁95) as computed with ISO 532-1 [58]. Whereas the 𝐿A,eq and 𝐿C,eq metrics are commonly used in noise
policies, the 𝑁95 was previously found to correlate strongly with the perceived loudness of traffic sounds [25].
For consistency and comparability, the scales for all items in the site evaluation questionnaire were normalized
such that all values ranged from −1 to 1 before further analysis was performed. The PAQ items were also transformed
into the normalized quantities “ISO Pleasantness (ISOPL)” and “ISO Eventfulness (ISOEV)” based on the definition
given in ISO 12913-3. Specifically, we computed ISOPL and ISOEV as
ISOPL =
2(𝑟pl −𝑟an)+√2(𝑟ca −𝑟ch +𝑟vi −𝑟mo)
8+8√2
∈ [−1,1],and (1)
ISOEV =2(𝑟ev −𝑟un)+√2(𝑟ch −𝑟ca +𝑟vi −𝑟mo)
8+8√2
∈ [−1,1],(2)
where 𝑟pl, 𝑟ev, 𝑟ch, 𝑟vi, 𝑟un , 𝑟ca, 𝑟an, 𝑟mo ∈ {1,2,…,5} are the extent to which the soundscape was respectively perceived
to be pleasant,eventful,chaotic,vibrant,uneventful,calm,annoying, and monotonous, on a scale of 1 to 5. Separate
positive affect (PA) and negative affect (NA) scores were also computed from the responses to the I-PANAS-SF, as
recommended by [59].
In the scope of a between-within experimental design, quantitative attributes were assessed using a two-way linear
mixed effects with a repeated measures approach. The factor within subjects, termed site, featured two levels: GND and
ROOF. Simultaneously, the between-subject factor, termed condition, featured two levels: AMB and AMSS.
For the examination of the attributes in DOMNoi,DOMNat,DOMHum,NA,OSQ,APPR,PLN}, a non-parametric
two-way linear mixed effects repeated measures type III rank-transformed analysis of variance (2ME-RT-RMANOVA)
was applied. This method involves replacing the original data with their ranks, a technique well suited for multiple
comparisons [60]. The model included a random intercept to account for potential variability in baseline responses
across participants.
We utilized a similar analytical approach to investigate the derived attributes in {PA,ISOPL,ISOEV,PRSSFas,
PRSSBA,PRSSCom,PRSSEC,PRSSES}, namely a non-parametric two-way linear mixed effects repeated measures type
III analysis of variance (2ME-RMANOVA). Notably, we refrained from rank transformation in this case, because the
residuals exhibited normality as confirmed through Shapiro-Wilk’s test (𝑝 > 0.05).
Lam et al.: Preprint submitted to Elsevier Page 12 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
In addition, to assess the potential impact of order effects and group sizes, multiple comparisons were made across
all the attributes of soundscape evaluation in {DOMNoi,DOMNat,DOMHum,NA,PA,OSQ,APPR,PLN,PRSSFas,
PRSSBA,PRSSCom,PRSSEC,PRSSES,ISOPL,ISOEV} for each condition, employing the non-parametric two-sample
Kolmogorov-Smirnov (KS) test. For the analysis of order effects, the responses were grouped into a sample from all
participants who evaluated the GND first followed by the ROOF, and another sample from all participants who evaluated
the ROOF first followed by the GND. For the analysis of group sizes, the responses were grouped into a sample from all
participants who evaluated the sites by themselves, and another sample from all participants who evaluated the sites
with at least one other participant in the same session. To mitigate false discovery rates due to multiple comparisons,
𝑝-values were adjusted using the Benjamini-Hochberg (BH) method separately for each condition.
All data analyses were conducted with the R programming language (Version 4.3.1; R Core Team [61]) on
a 64-bit ARM environment. Specifically, the analyses were performed with the following packages: KS test, BH
correction, Shapiro-Wilk Normality Test with stats (Version 4.3.1; R Core Team [61]); 2ME-RMANOVA and 2ME-
RT-RMANOVA with lmerTest (Version 3.1.3; Kuznetsova et al. [62]) and car (Version 3.1.2; Fox and Weisberg
[63]); Omega effect size with effectsize (Version 0.8.3; Ben-Shachar et al. [64]); and contrast tests with emmeans
(Version 1.8.7; Lenth [65]).
3. Results: Site evaluation questionnaire
A summary of the mean 𝜇and standard deviation 𝜎of quantities derived from the site evaluation questionnaire
is shown in Table 5. As mentioned in Section 2.6, all quantities are normalized to the same range [−1,1] for the
presentation of results in this section. Furthermore, only significant results of the 2ME-RT-RMANOVA and 2ME-
RMANOVA are presented here, but full details of the tests are given in Appendix B, Table B.1. For clarity, the scale
changes are illustrated in Figure 3, similarly organised by site and condition, where the significant posthoc contrast
pairs are accentuated.
3.1. Contrast by condition between groups at each site
At GND, no significant interaction effects were noted across the perceptual metrics between AMSS and AMB groups
(Figure 3, leftmost), aligning with expectations given the absence of AMSS at GND. Consistency in perception among
AMSS and AMB groups suggests stability in the GND soundscape and uniform participant perceptions, facilitating
comparison at ROOF.
At ROOF, the AMSS induced significant improvements in ISOPL,DOMNat,OSQ,PA,PRSSFas,PRSSBA, and
PRSSCom of the traffic-exposed soundscape (Figure 3, second from left). Notably, a 14.62 % increase in ISOPL marks
a key “positive transition” from a “bad” (𝜇= −0.19) to a “good” (𝜇= 0.10) soundscape, validating the efficacy of
Lam et al.: Preprint submitted to Elsevier Page 13 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table 5
Mean responses 𝜇(standard deviation 𝜎) of perceptual attributes in the site evaluation questionnaire investigated for the
validation study, organized by site and condition. The scales for all attributes are normalised to the range [−1,1]. Percentage
changes are computed between the AMB and AMSS for site, and between ROOF and GND for condition as scale changes on
the [−1,1] range with respect to the former. For instance, a change from −0.25 in the AMB condition to 0.75 in the AMSS
condition would be reported as a 50 % change. Significant changes as determined by posthoc tests are indicated in bold.
site condition
GND ROOF AMB AMSS
AMB AMSS Δ(%) AMB AMSS Δ(%) GND ROOF Δ(%) GND ROOF Δ(%)
DOMNoi 0.25 (0.44) 0.15 (0.50) -4.86 0.66 (0.39) 0.51 (0.42) -7.12 0.25 (0.44) 0.66 (0.39) 20.31 0.15 (0.50) 0.51 (0.42) 18.06
DOMNat 0.19 (0.40) 0.17 (0.49) -1.04 -0.36 (0.50) 0.19 (0.44) 27.69 0.19 (0.40) -0.36 (0.50) -27.34 0.17 (0.49) 0.19 (0.44) 1.39
DOMHum -0.25 (0.38) -0.24 (0.60) 0.69 -0.86 (0.34) -0.93 (0.34) -3.56 -0.25 (0.38) -0.86 (0.34) -30.47 -0.24 (0.60) -0.93 (0.34) -34.72
PA -0.07 (0.43) -0.07 (0.50) 0.12 -0.21 (0.38) 0.07 (0.59) 14.10 -0.07 (0.43) -0.21 (0.38) -7.03 -0.07 (0.50) 0.07 (0.59) 6.94
NA -0.88 (0.18) -0.88 (0.23) 0.17 -0.78 (0.30) -0.83 (0.43) -2.17 -0.88 (0.18) -0.78 (0.30) 4.84 -0.88 (0.23) -0.83 (0.43) 2.50
OSQ 0.17 (0.47) 0.14 (0.39) -1.65 -0.17 (0.50) 0.07 (0.55) 12.07 0.17 (0.47) -0.17 (0.50) -17.19 0.14 (0.39) 0.07 (0.55) -3.47
APPR -0.02 (0.39) 0.15 (0.44) 8.42 -0.38 (0.49) 0.01 (0.57) 19.44 -0.02 (0.39) -0.38 (0.49) -17.97 0.15 (0.44) 0.01 (0.57) -6.94
PLN -0.17 (0.35) -0.11 (0.49) 3.04 0.34 (0.43) 0.15 (0.55) -9.55 -0.17 (0.35) 0.34 (0.43) 25.78 -0.11 (0.49) 0.15 (0.55) 13.19
ISOPL 0.16 (0.32) 0.14 (0.30) -1.00 -0.19 (0.38) 0.10 (0.45) 14.62 0.16 (0.32) -0.19 (0.38) -17.47 0.14 (0.30) 0.10 (0.45) -1.86
ISOEV 0.03 (0.23) 0.05 (0.23) 1.19 0.06 (0.24) 0.08 (0.26) 1.18 0.03 (0.23) 0.06 (0.24) 1.50 0.05 (0.23) 0.08 (0.26) 1.49
PRSSFas -0.16 (0.44) -0.08 (0.39) 3.91 -0.49 (0.43) -0.06 (0.50) 21.22 -0.16 (0.44) -0.49 (0.43) -16.28 -0.08 (0.39) -0.06 (0.50) 1.04
PRSSBA 0.08 (0.59) 0.19 (0.48) 5.47 -0.22 (0.50) 0.30 (0.68) 25.97 0.08 (0.59) -0.22 (0.50) -14.71 0.19 (0.48) 0.30 (0.68) 5.79
PRSSCom -0.40 (0.35) -0.32 (0.30) 3.94 -0.66 (0.35) -0.38 (0.41) 13.72 -0.40 (0.35) -0.66 (0.35) -13.02 -0.32 (0.30) -0.38 (0.41) -3.24
PRSSEC -0.40 (0.33) -0.25 (0.29) 7.06 -0.61 (0.34) -0.35 (0.39) 13.24 -0.40 (0.33) -0.61 (0.34) -10.81 -0.25 (0.29) -0.35 (0.39) -4.63
PRSSES -0.34 (0.33) -0.28 (0.32) 3.04 -0.55 (0.29) -0.36 (0.31) 9.55 -0.34 (0.33) -0.55 (0.29) -10.68 -0.28 (0.32) -0.36 (0.31) -4.17
AMSS in improving PAQ. The ISOEV, a PAQ measure of a soundscape’s “Eventfulness” [13], was unaffected by the
AMSS intervention, as desired.
Though traffic noise dominance decreased insignificantly by 7.12%, a more-than-proportionate 27.69 % positive
transition in natural sound dominance from AMB (𝜇= −0.36) to AMSS (𝜇= 0.19) was observed at the ROOF. Significant
12.07 % positive transition from AMB (𝜇= −0.17) to AMSS (𝜇= 0.07) was also observed in OSQ, but increased
appropriateness of the soundscape (APPR; 19.44 %) and decreased perceived loudness (PLN;−9.55 %) were not
significant with the AMSS intervention at ROOF. AMSS significantly increase positive affect (PA; 14.10 %), but did not
significantly decrease negative affect (NA;−2.17 %).
The restorative potential of the AMSS was evidenced by significant improvements with AMSS at the ROOF in
PRSS dimensions of Fascination (21.22 %), Being-Away (25.97 %) and Compatibility (13.72 %). Particularly, a positive
transition was observed in PRSSBA from AMB (𝜇= −0.22) to AMSS (𝜇= 0.30), which is an indicator of respite
provided by the soundscape from daily stressors [40,41]. However, improvements in Extent sub-dimensions of PRSSEC
(13.24 %) and PRSSES (9.55 %) were not significant.
3.2. Contrast by sites within group under each condition
Under the AMB condition, which is indicative of the difference between the sites before intervention, significant
changes were noted in ISOPL,DOMNat,OSQ,PLN,PRSSFas,PRSSBA, and PRSSCom (Figure 3, second from right).
Lam et al.: Preprint submitted to Elsevier Page 14 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
GND
ROOF
AMB AMSS AMB AMSS
−1.0
−0.5
0.0
0.5
1.0
condition
Normalised mean
AMB
AMSS
GND ROOF GND ROOF
−1.0
−0.5
0.0
0.5
1.0
site
Normalised mean
Attribute
DOMNoi
DOMNat
DOMHum
PA
NA
OQ
APPR
PLN
ISOPL
ISOEV
PRSSFas
PRSSBA
PRSSCom
PRSSEC
PRSSES
Figure 3: Simple contrast of means across all perceptual attributes organized by condition and site. Contrasts by condition
are between group at each site, whereas contrasts by site are within group for each condition. The scales for all attributes
are normalised to the range [−1,1]. Significant differences as determined by posthoc contrast tests are accentuated
The PAQ in terms of ISOPL was rated a significant 17.47 % lower at the ROOF (𝜇= −0.19) than at the GND (𝜇= 0.16),
whereas ISOEV was equally neutral between the GND (𝜇= 0.03) and pre-intervention ROOF (𝜇= 0.06)sites.
As expected but not significant, traffic noise dominance was 20.31% higher at the traffic-exposed ROOF (𝜇= 0.66)
than at the GND (𝜇= 0.25). Additionally, human sounds were 30.47 % more dominant at GND (𝜇= −0.25) than at the
almost non-existent levels at the ROOF (𝜇= −0.86). On the other hand, natural sounds were scarce (𝜇= −0.36) and a
significant 27.34 % lower at ROOF than GND (𝜇= −0.19).
Before intervention, the OSQ was poor at the ROOF (𝜇= −0.17) and a significant 17.19% lower than the OSQ of the
GND (𝜇= 0.17). Similarly but not significantly, ROOF was rated 17.97 % less appropriate than the GND. Interestingly,
no significant changes in positive (PA) or negative affect (NA) scores were observed between GND and ROOF before
intervention.
Regarding restorative indicators, significant differences were noted in dimensions such as PRSSFas,PRSSBA, and
PRSSCom, indicating poorer restorativeness at the ROOF compared to GND. Notably, the restorativeness of GND was only
slightly conducive in terms of PRSSBA (𝜇= 0.08).
Under the AMSS condition, no significant changes were found between GND and ROOF sites, except for PLN (Figure 3,
rightmost). This suggests that the AMSS effectively improved the ISOPL,DOMNat,OSQ,PRSSFas,PRSSBA, and
Lam et al.: Preprint submitted to Elsevier Page 15 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table 6
Kendall correlation matrix between all attributes in the site evaluation questionnaire where the significance of each entry
in the upper triangle is denoted with a Holm-adjusted 𝑝-value and each entry in the lower triangle is denoted with an
unadjusted 𝑝-value. Asterisks indicate *𝑝 < 0.05; **𝑝 < 0.01; ***𝑝 < 0.001; ****𝑝 < 0.0001. The unit diagonal has been
removed for clarity.
DOMNoi DOM HumDOM Nat PA NA OSQ APPR PLN ISOPL ISOEV PRSS Fas PRSSBA PRSS ComPRSS EC PRSS ES
DOMNoi -.22 -.03 -.06 .06 **-.35 *-.30 ***.44 *-.30 .08 -.12 -.23 -.24 -.22 -.07
DOMHum *-.22 .27 .04 -.03 .08 .12 -.14 .09 .01 .12 .07 .07 .10 .17
DOMNat -.03 **.27 .18 -.11 *.29 .23 -.13 *.29 .02 ..28 .24 .25 *.30 *.29
PA -.06 .04 *.18 -.11 *.29 .23 -.02 .23 .01 **.35 ***.39 **.34 ***.37 **.35
NA .06 -.03 -.11 -.11 -.22 -.20 .16 *-.29 .04 -.07 -.10 -.21 -.14 -.00
OSQ ***-.35 .08 ***.29 ***.29 **-.22 ***.56 ***-.47 ***.62 -.13 **.34 ***.49 ***.54 ***.52 ..27
APPR ***-.30 .12 **.23 **.23 *-.20 ***.56 ***-.41 ***.50 -.03 **.35 ***.43 ***.48 ***.47 ..28
PLN ***.44 .-.14 -.13 -.02 ..16 ***-.47 ***-.41 ***-.39 .10 -.16 -.25 **-.35 *-.31 -.17
ISOPL ***-.30 .09 ***.29 **.23 ***-.29 ***.62 ***.50 ***-.39 -.03 **.34 ***.50 ***.52 ***.46 .25
ISOEV .08 .01 .02 .01 .04 -.13 -.03 .10 -.03 -.02 -.08 -.08 -.06 .03
PRSSFas -.12 .12 ***.28 ***.35 -.07 ***.34 ***.35 .-.16 ***.34 -.02 ***.61 ***.57 ***.55 ***.65
PRSSBA **-.23 .07 **.24 ***.39 -.10 ***.49 ***.43 **-.25 ***.50 -.08 ***.61 ***.69 ***.64 ***.51
PRSSCom **-.24 .07 **.25 ***.34 *-.21 ***.54 ***.48 ***-.35 ***.52 -.08 ***.57 ***.69 ***.65 ***.50
PRSSEC **-.22 .10 ***.30 ***.37 -.14 ***.52 ***.47 ***-.31 ***.46 -.06 ***.55 ***.64 ***.65 ***.52
PRSSES -.07 ..17 ***.29 ***.35 -.00 **.27 **.28 *-.17 **.25 .03 ***.65 ***.51 ***.50 ***.52
PRSSCom scores of the ROOF similar to those at the traffic-shielded GND. Although perceived loudness increased
(13.19 %), it was to a lesser extent than without AMSS intervention (25.78 %).
3.3. Correlation between subjective metrics
Based on the Holm-adjusted Kendall correlation, listed in Table 6,ISOPL was found to be significantly positively
correlated with DOMNat (𝜏= 0.2887), OSQ (𝜏= 0.6204), APPR (𝜏= 0.5048), and with the restorative metrics of
PRSSFas (𝜏= 0.3389), PRSSBA (𝜏= 0.4894), PRSSCom (𝜏= 0.5367), and PRSSEC (𝜏= 0.5215). In contrast, ISOPL
was negatively correlated with DOMNoi (𝜏= −0.3028), NA (𝜏= −0.2885), and PLN (𝜏= −0.3901), but was not
significantly associated with DOMHum,PA,ISOEV, and PRSSES.
3.4. Effect of order, group size and initial conditions
The KS tests with BH adjustments across each condition (AMSS and AMB) demonstrated that none of the attributes
from the site evaluation questionnaire were influenced by the order in which the participants assessed the sites
(GND→ROOF or ROOF→GND), as well as the number of participants in each session (1or >1). In other words, the results
of this study were not subject to potentially confounding order effects and the possibility of participants affecting each
others’ responses to the soundscapes experienced. Full details of the results can be found in Appendix B, Table B.2.
Posthoc contrast tests on PA between AMSS and AMB groups at the meeting point and the absence of interaction
effects on NA revealed no significant differences between the AMSS and AMB groups in terms of positive and negative
affect states before the commencement of the experiment.
Lam et al.: Preprint submitted to Elsevier Page 16 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
4. Results: Objective binaural measurements
At the ROOF, the mean 10-min 𝐿A,eq was 64.97± 3.38 dB(A) for the AMSS group and 63.96±2.95 dB(A) for the AMB
group, as shown in Table 7. Since AMSS was active for the AMSS group, it caused a slight but imperceptible increase
(about 1 dB(A)) in mean SPL over the study duration at the ROOF. This suggests that on average, AMSS selected
masker gains that were well below the ambient SPLs. For instance, if AMSS reproduced maskers at the same SPL as
the ambient acoustics, it would result in a 3dB(A) increase. This difference is further reduced to less than 1 dB(A)
when one of the AMSS sessions affected by aircraft noise was omitted from the computed mean.
At the GND, in contrast, the mean 10-min 𝐿A,eq was 63.78 ± 7.17 dB(A) for the AMSS group and 57.91 ± 1.46
dB(A) for the AMB group. The relatively higher mean SPL and standard deviation of SPL at the GFP for the AMSS
group was due to aircraft flybys occurring in three of sessions at the GND and one at ROOF, which when omitted from
the computation of the mean, would have given an 𝐿A,eq of 58.26 ± 1.77 dB(A) at the GND and 64.25 ± 1.07 dB(A) at
the ROOF instead. Hence, the difference in 𝐿A,eq between the GND and ROOF was about 6dB(A) in both AMB and AMSS
groups. A similar trend was observed in the C-weighted equivalent sound pressure level, 𝐿C,eq and in 𝑁95, where the
differences between the sites were about 3 to 5dB(C) and about 5 to 6 soneGF, respectively.
To examine the relationship between objective (psycho)acoustic parameters, and soundscape and restorative
indicators, a correlation and distribution analysis was conducted between objective parameters (𝐿A,eq,𝐿C,eq,𝑁95),
and soundscape (ISOPL,OSQ) and restorative (PRSSFas,PRSSBA,PRSSCom) indices that show statistical difference
between AMB–ROOF and AMB–GND in Section 3. The Holm-adjusted Kendall correlation revealed no significant
relationships between the (psycho)acoustic parameters and all the soundscape and restorative indices (Table B.3).
The disassociation between objective and perceptual indicators is further illustrated in the median contour plots
of the mean perceptual score for each session as a function of each (psycho)acoustic parameter, organised into the
condition–site pairs in Figure 4. Notably, distinct positive shift in median contours across all perceptual indicators was
achieved with the introduction of AMSS at the ROOF despite a similar levels of 𝐿A,eq,𝐿C,eq, or 𝑁95 in the AMB–ROOF
subgroup. Moreover, AMSS–ROOF exhibited similar ISOPL,PRSSFas,PRSSBA, and PRSSCom distributions as both GND
subgroups. Although the OSQ contours in AMSS–ROOF largely overlapped with AMB–ROOF across all objective indices,
there is a notable positive shift in the population distribution, as shown in Section 3.1. It is worth noting that the
𝐿C,eq distribution was greatly skewed by the dominant low-frequency content of aircraft flyby sounds in the AMSS–GND
sessions, which was not reflected in the 𝐿A,eq and 𝑁95.
Lam et al.: Preprint submitted to Elsevier Page 17 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table 7
Summary of mean 𝐿A,eq,𝐿C,eq,𝑁95 ,ISOPL,OSQ,PRSS Fas,PRSS BA , and PRSSCom values across 20 AMSS and 24 AMB
sessions in each of the GND and ROOF sites. Supplemented mean values for the AMSS sessions excluding aircraft flyby (3 in
GND; 1 in ROOF) are included.
Ambient AMSS AMSS (without aircraft flyby)
GND,N= 24 ROOF,N= 24 GND,N= 20 ROOF,N= 20 GND,N= 17 ROOF,N= 19
𝐿A,eq 57.91 (1.46) 63.96 (2.95) 61.04 (7.17) 64.97 (3.38) 58.26 (1.77) 64.25 (1.07)
𝐿C,eq 65.60 (1.55) 70.81 (2.54) 70.89 (6.42) 72.30 (3.27) 68.93 (4.39) 71.71 (2.01)
𝑁95 9.80 (0.87) 15.03 (1.64) 9.67 (0.31) 15.44 (0.87) 9.66 (0.34) 15.47 (0.88)
ISOPL 0.17 (0.32) −0.20 (0.37) 0.17 (0.23) 0.09 (0.38) 0.20 (0.23) 0.08 (0.38)
OSQ 0.17 (0.29) 0.03 (0.52) 0.14 (0.43) −0.22 (0.47) 0.21 (0.28) 0.00 (0.52)
PRSSFas −0.11 (0.38) −0.09 (0.44) −0.13 (0.44) −0.47 (0.44) −0.10 (0.40) −0.10 (0.45)
PRSSBA 0.21 (0.38) 0.32 (0.63) 0.10 (0.60) −0.21 (0.52) 0.25 (0.37) 0.30 (0.64)
PRSSCom −0.30 (0.24) −0.36 (0.36) −0.38 (0.32) −0.65 (0.37) −0.28 (0.24) −0.37 (0.37)
Lam et al.: Preprint submitted to Elsevier Page 18 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
50
55
60
65
70
50
55
60
65
70
50
55
60
65
70
50
55
60
65
70
50
55
60
65
70
LAeq, dB(A)
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
60
65
70
75
80
60
65
70
75
80
60
65
70
75
80
60
65
70
75
80
60
65
70
75
80
LCeq, dB(C)
ISOPL
OSQ
PRSSBA
PRSSCom
PRSSFas
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
−1.0 −0.5 0.0 0.5 1.0
5
10
15
20
5
10
15
20
5
10
15
20
5
10
15
20
5
10
15
20
N95, soneGF
site GND ROOF subgroup AMB−−GND AMB−−ROOF AMSS−−GND AMSS−−ROOF
Figure 4: Mean perceptual ISOPL,OSQ,PRSS Fas,PRSS BA , and PRSSCom scores across all participants per session (y-axis)
as a function of normalized objective 𝐿A,eq,𝐿C,eq, and 𝑁95 scores of each session (x-axis). Fifty percent of the sessions
lie within the median contours computed for AMB–GND,AMB–ROOF,AMSS–GND,AMSS–ROOF contrast subgroups. The left to
right columns represent 𝐿A,eq,𝐿C,eq , and 𝑁95, and each row represent each of the perceptual metrics, respectively.
Lam et al.: Preprint submitted to Elsevier Page 19 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
5. Discussion
For clarity, the research questions put forward in Section 1.4 are discussed sequentially in Section 5.1, Section 5.2,
and Section 5.3, respectively. The discussion culminates with the limitations of this study and suggestions for future
research in Section 5.4.
5.1. Assessing perceptual changes brought about by AMSS at the traffic-expose site
The lack of studies focusing on ISOPL as a design goal, especially in the context of augmenting soundscape affected
by traffic noise, highlights the novelty of our investigation. However, the findings could be placed in the context of a
previous virtual reality-based lab study set in a comparable scenario – an outdoor recreational space subjected to
traffic noise without direct visibility of the traffic source (i.e., location P2in [19]). In that study, the scale increase
in raw pleasantness at P2in [19] ranged from 5.00 to 18.33% across four types of bird sounds, and from −8.33 to
16.67 % across four types of water sounds, with each masker augmented 3dB(A) lower than the ambient traffic noise
levels at 65.2dB(A). With a higher increase in raw pleasantness of 23.35 % observed in this in-situ study (AMSS–ROOF:
𝑟𝑝𝑙 = 0.1389;AMB–ROOF:𝑟𝑝𝑙 = −0.3281), it is reasonable to conclude that the maskers selected by the AMSS indeed
prioritize maximizing ISOPL, where pleasantness is a significant component.
5.2. Perceptual implications of ISOPL as a soundscape intervention design goal
While the primary focus of AMSS optimization was ISOPL enhancement, significant improvements were evident
across various soundscape quality and restorative indicators. Notably, the consistent use of birdsongs as maskers led
to a significant increase in natural sound dominance (see Table 2), correlating with a reduction by 7.1% in DOMHum
and 3.6 % in DOMNoi, as explained by the informational masking theory [66,29].
With the modification of dominant sound source types, AMSS effectively enhanced the overall soundscape
quality (OSQ) at the traffic-exposed ROOF, surpassing the 9 % mean scale increase reported by [31] for their
manual augmentation approach. While caution is warranted in directly comparing methodologies due to differing
environments, the AMSS’s autonomous operation suggests a possible advantage over participant-led methods. Notably,
the OSQ contrast in the AMB condition between GND and ROOF, as described in Section 3.2, highlights the substantial
impact of traffic noise at the traffic-exposed ROOF. Additionally, the absence of significant differences between GND and
ROOF within the AMSS group suggests that AMSS could align the perception of OSQ at a traffic-exposed site with that
of the traffic-shielded environment.
The significant positive transition in positive affect (PA) induced by the AMSS suggests the potential for harnessing
the health benefits of natural sounds [14], which is made more accessible through its perception-driven autonomous
operation. It is important to note that non-optimized augmentation of natural sounds in urban environments could lead
to undesirable effects on mood and affect [67,68].
Lam et al.: Preprint submitted to Elsevier Page 20 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
On the contrary, the lack of significant changes in ISOEV suggests that AMSS did not alter the perceived
“Eventfulness” of the soundscape. This was likely due to the AMSS’s design goal, which focused solely on maximizing
ISOPL without affecting ISOEV. Additionally, according to the circumplex model of soundscape perception in ISO/TS
12913-2:2018, ISOPL and ISOEV are theoretically orthogonal axes, as observed by [69]. Thus, the absence of
significant differences in ISOEV serves as a validation of the circumplex model and underscores the efficacy of the
AMSS, which did not inadvertently impact ISOEV.
The AMSS intervention demonstrated its restorative potential through a significant increase in the Fascination,
Being-Away, and Compatibility PRSS dimensions. Particularly noteworthy was the 21.22% increase in PRSSFas,
indicating the maskers’ ability to captivate attention involuntarily [41], reinforcing the restorative effect of AMSS’s
informational masking mechanism [70]. Moreover, the significant shift of Being-Away (PRSSBA) from negative to
positive suggests AMSS effectively transformed the traffic-exposed ROOF soundscape from one associated with daily
stressors to a source of respite [41,40].
Nevertheless, while the rise in Compatibility (PRSSCom) due to AMSS was significant, its negative score fell
short of expectations afforded by natural soundscapes such as waterfronts or vast green spaces [70]. The restorative
limits of AMSS were also evident in both Extent-Coherence and Extent-Scope sub-dimensions. Despite a significant
increase in natural sound dominance, perceived coherency (PRSSEC) and expansiveness (PRSSES) of the environment
were unaffected, suggesting other factors, like visual impressions, may require adjustment. Notably, low PRSSEC and
PRSSES scores are characteristic of urban environments [70], consistent with observations in the GND. With significant
correlations between ISOPL and all PRSS dimensions except PRSSES (Extended Data Table 6), the positive link
suggests the potential for the AMSS to enhance PRSS alongside ISOPL, minimizing the need for separate models.
5.3. Impact of AMSS through (psycho)acoustic metrics and their relation with perceptual factors
The disconnection between (psycho)acoustic parameters and restorative indicators (i.e. PRSSFas,PRSSBA) contrasts
with [70], where 𝐿A,eq correlated negatively with both PRSSFas and PRSSBA, albeit with a brief 3 min stimuli exposure
time in [70]. This highlights the challenge of using objective metrics to assess the restorative impact of augmenting
“wanted” sounds in noisy environments.
Considering the GND had a mean 𝐿A,eq about 6 dB(A) lower than ROOF (Section 4), this suggests that AMSS
augmentation could be akin to a perceived 6 dB(A) noise reduction in terms of ISOPL. However, unlike previous
lab experiments [71], there was no change in PLN with AMSS, affirming that (psycho)acoustic parameters alone are
unable to fully capture soundscape perception changes.
Limitations of objective parameters in predicting subjective responses to soundscape augmentation were high-
lighted in an indoor experiment [25], where perceived annoyance was more accurately predicted by 𝐿C,eq and ISOPL
Lam et al.: Preprint submitted to Elsevier Page 21 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
than by objective parameters alone. Similarly, while 𝑁95 accurately predicted perceived traffic noise loudness indoors
[25], this did not hold true in this outdoor study, highlighting the need for caution in direct comparisons due to limited
data (sites and conditions).
5.4. Limitations and opportunities
The AI model in the AMSS used only acoustic data to determine the optimal masker-gain combinations [44].
However, factors such as participant demographics and visual environment could influence perception [72,73]. Future
AMSS versions could explore multimodal models, incorporating participant-linked information and real-time visual
data for broader applicability in different contexts [74,46]. Since ISOEV is orthogonal to ISOPL, future models could
optimize changes in ISOEV or a combination of both.
Drawing from an extensive survey and catalog of global soundscape interventions [75,76], the AMSS stands
out as the only AI-based intervention specifically engineered to autonomously elevate ISOPL levels. Notably, among
AI models trained on comprehensive datasets adhering to ISO 12913 standards [77,43,78], the AMSS hosts the sole
built-in prediction model capable of probabilistic modeling of ISOPL. The cloud-based framework behind AMSS could
potentially streamline soundscape interventions and monitoring on a large scale. It holds the key to cost-effective, large-
scale perceptual mapping compared to traditional methods reliant on human responses [79,80]. This advancement
could address challenges in widespread ISO 12913 standards adoption, particularly in predicting the socioeconomic
impact of soundscapes [81,80].
6. Conclusion
In conclusion, we described the implementation and validation of an AI-based soundscape augmentation system
(the AMSS) deployed at a pavilion at which road traffic was the dominant noise source in the acoustic environment.
Although the AMSS was designed only to select maskers for playback that maximized the ISOPL of the deployment
location, we found corresponding improvements in the rated overall quality, perceived restorativeness, appropriateness,
and positive affect by the participants in the validation study. The ISOPL of the deployment location was also found to
have increased to a level similar to that of a different pavilion where road traffic was significantly less dominant, and
where the objectively-measured SPL was signficantly lower. This was despite the fact that the AMSS caused a slight
increase in objectively-measured SPL at the deployment location due to the playback of maskers via a four-speaker
system.
In addition, the AMSS requires no human input to run, thereby allowing for reductions in time and labor required
to pick suitable maskers for augmentation as compared to traditional approaches involving expert guidance or post
hoc analysis of study results. The physical hardware of the AMSS was also installed after the pavilions had been
Lam et al.: Preprint submitted to Elsevier Page 22 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
built, with minimal alterations to the surrounding environment and infrastructure. Therefore, there is great potential to
further develop the AMSS and its corresponding soundscape augmentation approach for sustainable management of
noise pollution, especially in built-up areas where physical modifications to the surroundings to manage noise may be
impractical or unfeasible.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.
Acknowledgements
This research is supported by the Singapore Ministry of National Development and the National Research
Foundation, Prime Minister’s Office under the Cities of Tomorrow Research Programme (Award No. COT-V4-2020-1).
Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do
not reflect the view of National Research Foundation, Singapore, and Ministry of National Development, Singapore.
We would also like to thank the People’s Association for their support in the participant recruitment process, and the
Pasir Ris-Punggol Town Council for their assistance in the deployment of our system.
Part of this work was done while K. N. Watcharasupat was supported by the AAUW International Fellowship
from the American Association of University Women (AAUW), and separately by the IEEE Signal Processing Society
Scholarship.
Data Availability
The data that support the findings of this study are openly available in NTU research data repository DR-NTU
(Data) at https://doi.org/10.21979/N9/NEH5TR. The replication code used in this study is available on GitHub at the
following repository: https://doi.org/10.5281/zenodo.11141691. The code includes all the necessary scripts, functions,
and instructions to reproduce the results reported in the study.
CRediT authorship contribution statement
Bhan Lam: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Project admin-
istration, Data Curation, Writing - Original Draft, Writing - Review & Editing, Visualization, Supervision. Zhen-Ting
Ong: Project administration, Investigation, Data Curation. Kenneth Ooi: Methodology, Investigation, Software, Data
Curation, Validation, Formal analysis, Visualization, Writing - Original Draft, Writing - Review & Editing. Wen-Hui
Ong: Software, Data Curation. Trevor Wong: Software, Data Curation. Karn N. Watcharasupat: Software, Writing
Lam et al.: Preprint submitted to Elsevier Page 23 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
- Review & Editing. Vanessa Boey: Project administration, Investigation, Data Curation, Resources. Irene Lee:
Conceptualization, Writing - Review & Editing, Supervision, Resources, Funding acquisition. Joo Young Hong:
Methodology, Writing - Review & Editing, Supervision. Jian Kang: Methodology, Writing - Review & Editing,
Supervision. Kar Fye Alvin Lee: Formal analysis, Writing - Review & Editing. Georgios Christopoulos: Formal
analysis, Writing - Review & Editing, Funding acquisition. Woon-Seng Gan: Conceptualization, Writing - Review &
Editing, Supervision, Resources, Funding acquisition.
Lam et al.: Preprint submitted to Elsevier Page 24 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Appendix A. Questionnaires
Table A.1
Site evaluation questionnaire for the assessment of the soundscapes at the two study sites GND and ROOF. Participants
completed this questionnaire after a 10 min listening period at each site.
Question
Category
Instructions/Question Specific Items Rating Scale/Format
International
Positive and
Negative Affect
Schedule Short
Form
(I-PANAS-SF)
Indicate to what extent
you feel this way in this
moment.
Active Very slightly or not at
all–Extremely (5-point
categorical)
Attentive
Alert
Determined
Inspired
Hostile
Ashamed
Upset
Afraid
Nervous
Perceived Sound
Source
Dominance
(DOM)
To what extent do you
presently hear the
following types of
sounds?
Noise (e.g., traffic, construction, industry) Not at all–Dominates
completely (5-point
categorical scale)
Sounds from human beings (e.g.,
conversation, laughter, children at play,
footsteps)
Natural sounds (e.g., singing birds,
flowing water, wind in vegetation)
Perceived
Affective Quality
(PAQ)
For each of the 8 scales
below, to what extent
do you agree or disagree
that the surrounding
sound environment you
heard is ⋯
Eventful Strongly
disagree–Strongly agree
(5-point Categorical
scale)
Vibrant
Pleasant
Calm
Uneventful
Monotonous
Annoying
Chaotic
[Continued on next page]
Lam et al.: Preprint submitted to Elsevier Page 25 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Question
Category
Instructions/Question Specific Items Rating Scale/Format
Overall
Soundscape
Quality (OSQ)
Overall, how would you describe the present surrounding sound
environment?
Very good–Very bad
(5-point Categorical
scale)
Appropriateness
(APPR)
Overall, to what extent is the present surrounding sound
environment appropriate to the present place?
Not at all–Perfectly
(5-point Categorical
scale)
Perceived
Loudness (PLN)
How loud would you say the sound environment is? Not at all–Extremely
(5-point Categorical
scale)
Perceived
Restorativeness
Soundscape Scale
(PRSS) –
Fascination
(PRSSFas)
How much do you agree
with the following
statements?
My curiosity is awoken by these sounds Not at all–Completely
(7-point categorical
scale)
There are plenty of sounds for me to
discover
These sounds, I find fascinating
My interest is really held by following
what is going on with these sounds
PRSS –
Being-away
(PRSSBA)
How much do you agree
with the following
statements?
I get a break from my day-to-day routine
from spending time with these sounds
Not at all–Completely
(7-point categorical
scale)
I find that I don’t have to concentrate
much when I’m surrounded by these
sounds
The sounds give me a chance to step back
from things that demand my focus
I feel free from work and/or
responsibilities when I am with these
sounds
These sounds are a refuge for me from
unwanted distractions
[Continued on next page]
Lam et al.: Preprint submitted to Elsevier Page 26 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Question
Category
Instructions/Question Specific Items Rating Scale/Format
PRSS –
Compatibility
(PRSSCom)
How much do you agree
with the following
statements?
I rapidly adapt to these sounds Not at all–Completely
(7-point categorical
scale)While I am with these sounds, it is easy to
do what I want
The sounds fit well with my preferences
PRSS –
Extent-Coherence
(PRSSEC)
How much do you agree
with the following
statements?
The existing sounds belong to this
soundscape
Not at all–Completely
(7-point categorical
scale)The sounds blend together to create a
harmonious soundscape
The sounds in this environment are
well-organized, which makes it easy for me
to hear the relationships between them
PRSS –
Extent-Scope
(PRSSES)
How much do you agree
with the following
statements?
There are lots of different sounds to
explore in this place
Not at all–Completely
(7-point categorical
scale)The sounds make it feel like this place is
vast
These sounds have the quality to create a
world of their own
Lam et al.: Preprint submitted to Elsevier Page 27 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table A.2
Derviation of the Perceived Restorative Soundscape Scale (PRSS) items
PRSS
Dimensions
PRSS Items (specific framing in Payne and
Guastavino [41])
PRSS Items (this study) Remarks
Fascination My curiosity is awoken by these sounds -
There are plenty of sounds for me to discover -
These sounds, I find fascinating -
My interest is really held by following what is going on with these sounds -
Being-Away I get a break from my day-to-day routine from spending time with these sounds -
My concentration is demanded by these
sounds
I find that I don’t have to concentrate much
when I’m surrounded by these sounds
rephrased
From these sounds, I experience few
attentional demands
The sounds give me a chance to step back
from things that demand my focus
rephrased
I feel free from work and/or responsibilities when I am with these sounds -
I need to think of my obligations when I am
with these sounds
-removed
These sounds are a refuge for me from unwanted distractions
Compatibility There is an accordance between these sounds
and what I like to do
- removed
I rapidly adapt to these sounds -
While I am with these sounds, it is easy to do what I want -
My personal inclinations fits with being with
these sounds
The sounds fit well with my preferences rephrased
Extent
(Coherence)
The existing sounds belong to this soundscape -
The sounds fit together to form a coherent
soundscape
The sounds blend together to create a
harmonious soundscape
rephrased
These sounds are coherent The sounds in this environment are
well-organized, which makes it easy for me to
hear the relationships between them
combined
The sounds are clearly organized
The physical arrangement of these sounds has
a clear order
Extent (Scope) There are plenty of sounds to allow
exploration in many directions
There are lots of different sounds to explore
in this place
rephrased
The extent of these sounds seems limitless The sounds make it feel like this place is vast combined
These sounds feel very spacious
These sounds have the quality of being a
whole world to themselves
These sounds have the quality to create a
world of their own
rephrased
Lam et al.: Preprint submitted to Elsevier Page 28 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table A.3
Participant information questionnaire administered prior to the end of each session. Participants completed this
questionnaire after the soundscape evaluations had been completed at both study sites RTGP and GFP.
Question
Category
Instructions/Questions Specific Items Rating Scale/Format
Gender What is your gender? Male/Female/Non-
conforming/Prefer not
to say
Age What is your age? Integer in [21,70]
Occupation What is your occupational status? Employed/Unemployed/
Retired/Student/Rather
not say/Other
Individual Noise
Sensitivity (INS)
Select the option that
best represents your
level of agreement with
the statement.
I wouldn’t mind living on a noisy street if
the apartment I had was nice.
Strongly
disagree–Strongly agree
(5-point categorical
scale)
I am more aware of noise than I used to
be.
No one should mind much if someone
turns up his stereo full blast once in a
while.
At movies, whispering and crinkling candy
wrappers disturb me.
I am easily awakened by noise.
If it’s noisy where I’m studying, I try to
close the door or window or move
someplace else.
I get annoyed when my neighbors are
noisy.
I get used to most noises without much
difficulty.
How much would it matter to you if an
apartment you were interested in renting
was located across from a fire station?
[Continued on next page]
Lam et al.: Preprint submitted to Elsevier Page 29 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Question
Category
Instructions/Questions Specific Items Rating Scale/Format
Sometimes noises get on my nerves and
get me irritated.
Strongly
disagree–Strongly agree
(5-point categorical
scale)
Even music I normally like will bother me
if I’m trying to concentrate.
It wouldn’t bother me to hear the sounds
of everyday living from neighbors
(footsteps, running water, etc).
When I want to be alone, it disturbs me
to hear outside noises.
I’m good at concentrating no matter what
is going on around me.
In a library, I don’t mind if people carry on
a conversation if they do it quietly.
There are often times when I want
complete silence.
Motorcycles ought to be required to have
bigger mufflers.
I find it hard to relax in a place that’s
noisy.
I get mad at people who make noise that
keeps me from falling asleep or getting
work done.
I wouldn’t mind living in an apartment
with thin walls.
I am sensitive to noise.
Baseline Noise
Annoyance
(BNA)
In general, how much
does noise from
bother,
disturb, or annoy you?
Aircraft (military or civilian) Not at all–Extremely
(5-point Categorical
scale)
Road traffic
MRT (trains)
Children
Other people
Animals
[Continued on next page]
Lam et al.: Preprint submitted to Elsevier Page 30 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Question
Category
Instructions/Questions Specific Items Rating Scale/Format
Construction worksites
Construction (renovations)
Any other noises
Perceived Stress
Scale (PSS)
In the last month, how
often have you...
been upset because of something that
happened unexpectedly?
Never–Very often
(5-point categorical
scale)felt that you were unable to control the
important things in your life?
felt nervous and “stressed”?
felt confident about your ability to handle
your personal problems?
felt that things were going your way?
found that you could not cope with all the
things that you had to do?
been able to control irritations in your life?
felt that you were on top of things?
been angered because of things that were
outside of your control?
felt difficulties were piling up so high that
you could not overcome them?
WHO-Five
Well-Being Index
(WHO-5)
For each of these
statements, which is the
closest to how you have
been feeling over the
last two weeks?
I have felt cheerful and in good spirits. At no time–All of the
time (6-point
Categorical scale)
I have felt calm and relaxed.
I have felt active and vigorous.
I woke up feeling fresh and rested.
My daily life has been filled with things
that interest me.
Lam et al.: Preprint submitted to Elsevier Page 31 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Appendix B. Statistical results
Table B.1: Summary of statistical tests for attributes in soundscape evaluation questionnaire (sound source dominance,
overall quality, appropriateness, loudness, ISOPL,ISOEV, and PRSS dimensions) across site (GND and ROOF), condition
(AMSS and AMB), and their interaction (site:condition). Test abbreviations and symbols for significance levels and effect
sizes are defined in the footnote.
Term Test1𝑝−value2Effect Size3
Sound source dominance – Noise (DOMNoi)
site 2ME-RT-RMANOVA ****0.0000 (L)0.3182
condition 2ME-RT-RMANOVA 0.1571 (S)0.0145
site:condition 2ME-RT-RMANOVA 0.5667 0.0000
Sound source dominance – Natural sounds (DOMNat)
site 2ME-RT-RMANOVA ***0.0004 (L)0.1464
condition 2ME-RT-RMANOVA **0.0015 (M)0.1175
site:condition 2ME-RT-RMANOVA ***0.0003 (L)0.1492
ÂăÂăÂăÂăAMB–AMSS |GND Simple Contrasts for Condition 0.9513 (S)0.0149
ÂăÂăÂăÂăAMB–AMSS |ROOF Simple Contrasts for Condition ****0.0000 (L)-1.1574
ÂăÂăÂăÂăGND–ROOF |AMB Simple Contrasts for Site ****0.0000 (L)1.1661
ÂăÂăÂăÂăGND–ROOF |AMSS Simple Contrasts for Site 0.9783 -0.0061
Sound source dominance – Human sounds (DOMHum)
site 2ME-RT-RMANOVA ****0.0000 (L)0.5180
condition 2ME-RT-RMANOVA 0.1039 (S)0.0121
site:condition 2ME-RT-RMANOVA 0.8785 0.0000
Positive Affect (PA)
Residuals Shapiro-Wilk normality test 0.1731 -
site 2ME-RMANOVA 0.6753 0.0000
condition 2ME-RMANOVA 0.1620 (S)0.0138
site:condition 2ME-RMANOVA *0.0211 (S)0.0403
ÂăÂăÂăÂăAMB–AMSS |GND Simple Contrasts for Condition 0.9835 -0.0050
ÂăÂăÂăÂăAMB–AMSS |MP Simple Contrasts for Condition 0.2242 (L)-0.2963
ÂăÂăÂăÂăAMB–AMSS |ROOF Simple Contrasts for Condition *0.0179 (L)-0.5839
ÂăÂăÂăÂăGND–MP|AMB Simple Contrasts for Site 0.8971 (S)0.0669
ÂăÂăÂăÂăGND–ROOF |AMB Simple Contrasts for Site 0.1369 (L)0.2912
ÂăÂăÂăÂăMP–ROOF |AMB Simple Contrasts for Site 0.2999 (L)0.2242
ÂăÂăÂăÂăGND–MP|AMSS Simple Contrasts for Site 0.2625 (L)-0.2243
ÂăÂăÂăÂăGND–ROOF |AMSS Simple Contrasts for Site 0.1133 (L)-0.2876
ÂăÂăÂăÂăMP–ROOF |AMSS Simple Contrasts for Site 0.8977 (S)-0.0633
Continues to the next page...
Lam et al.: Preprint submitted to Elsevier Page 32 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas Continued from the previous page...
Term Test1𝑝−value2Effect Size3
Negative Affect (NA)
Residuals Shapiro-Wilk normality test ****0.0000 -
site 2ME-RT-RMANOVA 0.3525 0.0006
condition 2ME-RT-RMANOVA *0.0253 (S)0.0550
site:condition 2ME-RT-RMANOVA 0.1665 (S)0.0114
Overall soundscape quality (OSQ)
site 2ME-RT-RMANOVA **0.0041 (M)0.0965
condition 2ME-RT-RMANOVA 0.2204 0.0073
site:condition 2ME-RT-RMANOVA *0.0271 (S)0.0540
ÂăÂăÂăÂăAMB–AMSS |GND Simple Contrasts for Condition 0.7087 (M)0.0910
ÂăÂăÂăÂăAMB–AMSS |ROOF Simple Contrasts for Condition *0.0221 (L)-0.5631
ÂăÂăÂăÂăGND–ROOF |AMB Simple Contrasts for Site ***0.0009 (L)0.7525
ÂăÂăÂăÂăGND–ROOF |AMSS Simple Contrasts for Site 0.6297 (M)0.0984
Appropriateness (APPR)
site 2ME-RT-RMANOVA **0.0024 (M)0.1074
condition 2ME-RT-RMANOVA ***0.0007 (M)0.1327
site:condition 2ME-RT-RMANOVA 0.1591 (S)0.0142
Perceived loudness (PLN)
site 2ME-RT-RMANOVA ****0.0000 (L)0.3561
condition 2ME-RT-RMANOVA 0.5667 0.0000
site:condition 2ME-RT-RMANOVA *0.0221 (S)0.0587
ÂăÂăÂăÂăAMB–AMSS |GND Simple Contrasts for Condition 0.4189 (L)-0.1971
ÂăÂăÂăÂăAMB–AMSS |ROOF Simple Contrasts for Condition .0.0812 (L)0.4274
ÂăÂăÂăÂăGND–ROOF |AMB Simple Contrasts for Site ****0.0000 (L)-1.1600
ÂăÂăÂăÂăGND–ROOF |AMSS Simple Contrasts for Site **0.0057 (L)-0.5355
Continues to the next page...
Lam et al.: Preprint submitted to Elsevier Page 33 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas Continued from the previous page...
Term Test1𝑝−value2Effect Size3
ISO Pleasantness (ISOPL)
Residuals Shapiro-Wilk normality test 0.1229 -
site 2ME-RMANOVA **0.0011 (M)0.1248
condition 2ME-RMANOVA *0.0432 (S)0.0434
site:condition 2ME-RMANOVA **0.0082 (M)0.0808
ÂăÂăÂăÂăAMB–AMSS |GND Simple Contrasts for Condition 0.8241 (S)0.0541
ÂăÂăÂăÂăAMB–AMSS |ROOF Simple Contrasts for Condition **0.0014 (L)-0.7926
ÂăÂăÂăÂăGND–ROOF |AMB Simple Contrasts for Site ***0.0001 (L)0.9473
ÂăÂăÂăÂăGND–ROOF |AMSS Simple Contrasts for Site 0.6487 (M)0.1006
ISO Eventfulness (ISOEV)
Residuals Shapiro-Wilk normality test 0.7790 -
site 2ME-RMANOVA 0.4576 0.0000
condition 2ME-RMANOVA 0.5795 0.0000
site:condition 2ME-RMANOVA 0.9990 0.0000
Perceived Restorativeness Soundscape Scale: Fascination (PRSSFas)
Residuals Shapiro-Wilk normality test 0.8728 -
site 2ME-RMANOVA *0.0203 (M)0.0606
condition 2ME-RMANOVA **0.0034 (M)0.1000
site:condition 2ME-RMANOVA **0.0083 (M)0.0806
ÂăÂăÂăÂăAMB–AMSS |GND Simple Contrasts for Condition 0.4713 (L)-0.1755
ÂăÂăÂăÂăAMB–AMSS |ROOF Simple Contrasts for Condition ***0.0001 (L)-0.9538
ÂăÂăÂăÂăGND–ROOF |AMB Simple Contrasts for Site **0.0011 (L)0.7314
ÂăÂăÂăÂăGND–ROOF |AMSS Simple Contrasts for Site 0.8178 (M)-0.0468
Perceived Restorativeness Soundscape Scale: Being-Away (PRSSBA)
Residuals Shapiro-Wilk normality test 0.7777 -
site 2ME-RMANOVA 0.3081 0.0006
condition 2ME-RMANOVA **0.0034 (M)0.1005
site:condition 2ME-RMANOVA *0.0193 (M)0.0618
ÂăÂăÂăÂăAMB–AMSS |GND Simple Contrasts for Condition 0.4309 (L)-0.1920
ÂăÂăÂăÂăAMB–AMSS |ROOF Simple Contrasts for Condition ***0.0003 (L)-0.9116
ÂăÂăÂăÂăGND–ROOF |AMB Simple Contrasts for Site *0.0241 (L)0.5165
ÂăÂăÂăÂăGND–ROOF |AMSS Simple Contrasts for Site 0.3390 (L)-0.2031
Continues to the next page...
Lam et al.: Preprint submitted to Elsevier Page 34 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas Continued from the previous page...
Term Test1𝑝−value2Effect Size3
Perceived Restorativeness Soundscape Scale: Compatibility (PRSSCom)
Residuals Shapiro-Wilk normality test 0.3328 -
site 2ME-RMANOVA ***0.0009 (M)0.1287
condition 2ME-RMANOVA *0.0135 (M)0.0698
site:condition 2ME-RMANOVA *0.0456 (S)0.0422
ÂăÂăÂăÂăAMB–AMSS |GND Simple Contrasts for Condition 0.3652 (L)-0.2209
ÂăÂăÂăÂăAMB–AMSS |ROOF Simple Contrasts for Condition **0.0020 (L)-0.7697
ÂăÂăÂăÂăGND–ROOF |AMB Simple Contrasts for Site ***0.0005 (L)0.7308
ÂăÂăÂăÂăGND–ROOF |AMSS Simple Contrasts for Site 0.3378 (L)0.1819
Perceived Restorativeness Soundscape Scale: Extent-Coherence (PRSSEC)
Residuals Shapiro-Wilk normality test 0.9051 -
site 2ME-RMANOVA **0.0015 (M)0.1182
condition 2ME-RMANOVA **0.0023 (M)0.1089
site:condition 2ME-RMANOVA 0.2031 0.0090
Perceived Restorativeness Soundscape Scale: Extent-Scope (PRSSES)
Residuals Shapiro-Wilk normality test 0.0581 -
site 2ME-RMANOVA **0.0010 (M)0.1254
condition 2ME-RMANOVA *0.0410 (S)0.0446
site:condition 2ME-RMANOVA 0.1504 (S)0.0155
1Two-way linear mixed effects repeated measures Type III ANOVA (2ME-RMANOVA); Two-way linear mixed effects repeated measures
Type III Rank-transformed ANOVA (2ME-RT-RMANOVA)
2*𝑝 < 0.05; **𝑝 < 0.01; ***𝑝 < 0.001; ****𝑝 < 0.0001
3Partial Omega squared (𝜔2
𝑝) for linear mixed effects and Cohen’s 𝑑for simple contrasts. (L) large effect >0.14 ; (M) medium effect
>0.06; (S) small effect >0.01
Lam et al.: Preprint submitted to Elsevier Page 35 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
Table B.2
Summary of exact two-sample Kolmogorov-Smirnov tests to examine effect of order (GND–ROOF or ROOF–GND) and group
size (1 or >1) on each soundscape evaluation attribute (sound source dominance, overall quality, appropriateness, loudness,
ISOPL,ISOEV, and PRSS dimensions) across each condition (AMSS and AMB). All the 𝑝-values were adjusted for multiple
comparisons within conditions with the Benjamini-Hochberg (BH) method.
DOMNoi DOM Nat DOMHum OSQ APPR PLN ISOPL ISOEV PA NA PRSSFas PRSS BA PRSS ComPRSSEC PRSS ES
Order
AMB 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
AMSS 0.95 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83
Group Size
AMB 0.98 0.98 0.98 0.98 0.96 0.96 0.98 0.96 0.98 0.96 0.98 0.96 0.96 0.98 0.96
AMSS 0.75 0.75 0.94 0.75 0.75 0.94 0.75 0.94 0.75 0.94 0.80 0.80 0.75 0.75 0.80
Table B.3
Kendall correlation matrix between all objective acoustic measures and perceptual attributes in the site evaluation
questionnaire where the significance of each entry in the upper triangle is denoted with a Holm-adjusted 𝑝-value and each
entry in the lower triangle is denoted with an unadjusted 𝑝-value. Asterisks indicate *𝑝 < 0.05; **𝑝 < 0.01; ***𝑝 < 0.001;
****𝑝 < 0.0001. The unit diagonal has been removed for clarity.
ISOPL OSQ PA PLN PRSSFas PRSS BA PRSS Com 𝐿A,eq 𝐿C,eq 𝑁95
ISOPL ***0.64 0.31 **-0.40 **0.40 ***0.56 ***0.61 -0.22 -0.10 -0.18
OSQ ***0.64 0.32 ***-0.45 0.29 ***0.49 ***0.52 -0.19 -0.09 -0.16
PA **0.31 **0.32 -0.05 **0.40 ***0.44 **0.40 0.00 -0.04 0.04
PLN ***-0.40 ***-0.45 -0.05 -0.20 -0.31 ***-0.44 0.29 0.25 0.28
PRSSFas ***0.40 **0.29 ***0.40 -0.20 ***0.59 ***0.59 -0.09 -0.07 -0.11
PRSSBA ***0.56 ***0.49 ***0.44 **-0.31 ***0.59 ***0.71 -0.05 0.01 -0.05
PRSSCom ***0.61 ***0.52 ***0.40 ***-0.44 ***0.59 ***0.71 -0.19 -0.12 -0.18
𝐿A,eq *-0.22 .-0.19 0.00 **0.29 -0.09 -0.05 -0.19 ***0.59 ***0.68
𝐿C,eq -0.10 -0.09 -0.04 *0.25 -0.07 0.01 -0.12 ***0.59 ***0.47
𝑁95 -0.18 -0.16 0.04 **0.28 -0.11 -0.05 -0.18 ***0.68 ***0.47
Lam et al.: Preprint submitted to Elsevier Page 36 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
References
[1] World Health Organization Regional Office for Europe, Environ-
mental Noise Guidelines for the European Region, The Regional
Office for Europe of the World Health Organization, Copenhagen,
Denmark, 2018.
[2] Department of Environment Food and Rural Affairs,
Noise Pollution: Economic Analysis, 2014. URL: https:
//www.gov.uk/guidance/noise-pollution-economic-analysis#
full-publication-update-history.
[3] E. A. King, Environment: Science and Policy for Sustainable
Development 64 (2022) 17–32. doi:10.1080/00139157.2022.
2046456, publisher: Routledge.
[4] D. Fink, Acoustics Today 15 (2019) 38. doi:10.1121/AT.2019.
15.3.38.
[5] J. B. Newbury, J. Heron, J. B. Kirkbride, H. L. Fisher, I. Bakolis,
A. Boyd, R. Thomas, S. Zammit, JAMA Network Open 7 (2024)
e2412169. URL: https://doi.org/10.1001/jamanetworkopen.2024.
12169. doi:10.1001/jamanetworkopen.2024.12169.
[6] O. Hahad, M. Kuntic, S. Al-Kindi, I. Kuntic, D. Gilan,
K. Petrowski, A. Daiber, T. MÃijnzel, Journal of Exposure
Science & Environmental Epidemiology (2024) 1–8.
URL: https://www.nature.com/articles/s41370-024-00642-5.
doi:10.1038/s41370-024- 00642-5, publisher: Nature
Publishing Group.
[7] R. Guski, D. Schreckenberg, R. Schuemer, International Journal
of Environmental Research and Public Health 14 (2017) 1–39.
doi:10.3390/ijerph14121539.
[8] J. Kang, Urban Sound Environment, Taylor & Francis, London,
UK, 2007.
[9] J. Kang, Frontiers of Engineering Management 4 (2017)
184–194. URL: http://engineering.cae.cn/fem/EN/10.15302/
J-FEM-2017026. doi:10.15302/J-FEM- 2017026.
[10] J. Kang, F. Aletta, T. T. Gjestland, L. A. Brown, D. Bottel-
dooren, B. Schulte-Fortkamp, P. Lercher, I. van Kamp, K. Genuit,
A. Fiebig, J. L. Bento Coelho, L. Maffei, L. Lavia, Building
and Environment 108 (2016) 284–294. URL: http://linkinghub.
elsevier.com/retrieve/pii/S0360132316303067. doi:10.1016/j.
buildenv.2016.08.011, iSBN: 03601323.
[11] International Organization for Standardization, ISO 12913-1:2014
Acoustics âĂŤ Soundscape âĂŤ Part 1 : Definition and concep-
tual framework, International Organization for Standardization,
Geneva, Switzerland, 2014.
[12] International Organization for Standardization, ISO/TS 12913-
2:2018 Acoustics âĂŤ Soundscape âĂŤ Part 2: Data collection
and reporting requirements, International Organization for Stan-
dardization, Geneva, Switzerland, Switzerland, 2018.
[13] International Organization for Standardization, ISO/TS 12913-
3:2019 - Acoustics âĂŤ Soundscape - Part 3: Data analysis, In-
ternational Organization for Standardization, Geneva, Switzerland,
2019.
[14] R. T. Buxton, A. L. Pearson, C. Allou, K. Fristrup, G. Wit-
temyer, Proceedings of the National Academy of Sciences 118
(2021) e2013097118. URL: https://www.pnas.org/content/118/14/
e2013097118. doi:10.1073/pnas.2013097118, publisher: Na-
tional Academy of Sciences.
[15] United Nations Environment Programme, Frontiers 2022: Noise,
Blazes and Mismatches âĂŞ Emerging Issues of Environmental
Concern, Nairobi, Kenya, 2022.
[16] J. C. Fisher, M. Dallimer, K. N. Irvine, S. G. Aizlewood,
G. E. Austen, R. D. Fish, P. M. King, Z. G. Davies, Nature
Sustainability 6 (2023) 1219–1227. URL: well. doi:10.1038/
s41893-023- 01151-3, publisher: Springer US.
[17] J. Y. Jeon, P. J. Lee, J. You, J. Kang, The Journal of the Acousti-
cal Society of America 127 (2010) 1357–1366. doi:10.1121/1.
3298437, iSBN: 0001-4966.
[18] Y. Hao, J. Kang, H. WÃűrtche, The Journal of the Acoustical So-
ciety of America 140 (2016) 978–987. doi:10.1121/1.4960570,
publisher: Acoustical Society of America.
[19] J. Y. Hong, B. Lam, Z.-T. Ong, K. Ooi, W.-S. Gan, J. Kang,
S. Yeong, I. Lee, S.-T. Tan, Sustainable Cities and Society 63
(2020) 102475. doi:10.1016/j.scs.2020.102475, publisher:
Elsevier Ltd.
[20] J. Y. Hong, B. Lam, Z.-T. Ong, R. Gupta, W.-S. Gan, Proceedings
of the 24th International Congress on Sound and Vibration (2017)
1–6.
Lam et al.: Preprint submitted to Elsevier Page 37 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
[21] W. Yang, H. J. Moon, Applied Acoustics 145 (2019) 234–
244. doi:10.1016/j.apacoust.2018.10.015, publisher: Else-
vier Ltd.
[22] J. Y. Jeon, P. J. Lee, J. You, J. Kang, The Journal of the Acousti-
cal Society of America 131 (2012) 2101–2109. doi:10.1121/1.
3681938, iSBN: 1520-8524.
[23] M. RÃědsten-Ekman, Ã. Axelsson, M. E. Nilsson, Acta
Acustica united with Acustica 99 (2013) 218–225. URL:
http://openurl.ingenta.com/content/xref?genre=article&
issn=1610-1928&volume=99&issue=2&spage=218.
doi:10.3813/AAA.918605.
[24] J. You, P. J. Lee, J. Y. Jeon, Noise Control Engineering Journal 58
(2010) 477. doi:10.3397/1.3484183.
[25] B. Lam, K. C. Q. Lim, K. Ooi, Z.-T. Ong, D. Shi, W.-S. Gan,
Sustainable Cities and Society (2023) 104763. doi:10.1016/j.
scs.2023.104763.
[26] G. CerwÃľn, Landscape Research 41 (2016) 481–494. doi:10.
1080/01426397.2015.1117062.
[27] F. M. Calarco, L. Galbrun, Applied Acoustics 219 (2024)
109947. URL: https://doi.org/10.1016/j.apacoust.2024.109947.
doi:10.1016/j.apacoust.2024.109947, publisher: Elsevier
Ltd.
[28] M. E. Nilsson, J. Alvarsson, M. RÃědsten-Ekman, K. Bolin,
Noise Control Engineering Journal 58 (2010) 524. URL:
http://www.ingentaconnect.com/content/ince/ncej/2010/
00000058/00000005/art00007. doi:10.3397/1.3484182.
[29] C. K. Chau, T. M. Leung, W. K. Chung, S. K. Tang, Applied Acous-
tics 213 (2023) 109650. URL: https://doi.org/10.1016/j.apacoust.
2023.109650. doi:10.1016/j.apacoust.2023.109650, pub-
lisher: Elsevier Ltd.
[30] M. Hedblom, B. Gunnarsson, M. Schaefer, I. Knez,
P. Thorsson, J. N. LundstrÃűm, International Journal of
Environmental Research and Public Health 16 (2019).
URL: https://www.proquest.com/scholarly-journals/
sounds-nature-city-no-evidence-bird-song/docview/
2329226373/se-2?accountid=12691NS-. doi:10.3390/
ijerph16081390, publisher: MDPI AG Place: Basel ISBN:
4618671041.
[31] T. Van Renterghem, K. Vanhecke, K. Filipan, K. Sun,
T. De Pessemier, B. De Coensel, W. Joseph, D. Botteldooren,
Landscape and Urban Planning 194 (2020) 103705. URL:
https://linkinghub.elsevier.com/retrieve/pii/S016920461931093X.
doi:10.1016/j.landurbplan.2019.103705, publisher:
Elsevier.
[32] B. Lam, W.-S. Gan, D. Shi, M. Nishimura, S. Elliott, Building and
Environment 200 (2021) 107928. doi:10.1016/j.buildenv.
2021.107928.
[33] J. Y. Hong, B. Lam, Z.-T. Ong, K. Ooi, W.-S. Gan, J. Kang,
S. Yeong, I. Lee, S.-T. Tan, Building and Environment 194
(2021) 107688. URL: https://linkinghub.elsevier.com/retrieve/
pii/S0360132321000998. doi:10.1016/j.buildenv.2021.
107688, publisher: Pergamon.
[34] R. Regazzi, B. Cunha, H. V. d. Miranda, J. J. GÃşmez Acosta,
C. R. Hall Barbosa, M. N. Frota, J. V. Souza, C. A.
Machado Gomes, Applied Sciences 11 (2021) 7771.
URL: https://www.proquest.com/scholarly-journals/
development-validation-masking-system-mitigation/docview/
2570578969/se-2NS-. doi:http://dx.doi.org/10.3390/
app11177771, publisher: MDPI AG Place: Basel.
[35] L. Lenne, P. Chevret, J. Marchand, Applied Acoustics 158
(2020) 107049. doi:10.1016/j.apacoust.2019.107049, pub-
lisher: Elsevier Ltd.
[36] R. Nordahl, Eurasip Journal on Audio, Speech, and
Music Processing 2010 (2010). URL: https://www.
scopus.com/inward/record.uri?eid=2-s2.0-79251564887&
doi=10.1155%2F2010%2F426937&partnerID=40&md5=
305958c69a3584a26bc1547ee67c4d1dNS-. doi:10.1155/2010/
426937.
[37] R. Nordahl, L. Turchet, S. Serafin, IEEE Transactions on
Visualization and Computer Graphics 17 (2011) 1234–
1244. URL: https://www.proquest.com/scholarly-journals/
sound-synthesis-evaluation-interactive-footsteps/docview/
876217463/se-2?accountid=12691NS-. doi:http://dx.doi.
org/10.1109/TVCG.2011.30, publisher: The Institute of
Electrical and Electronics Engineers, Inc. (IEEE) Place: New
York.
Lam et al.: Preprint submitted to Elsevier Page 38 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
[38] L. Turchet, S. Serafin, Appl Acoust 74 (2013) 566–574.
URL: https://www.scopus.com/inward/record.uri?eid=2-s2.
0-84878053328&doi=10.1016%2Fj.apacoust.2012.10.010&
partnerID=40&md5=e871f6f552e7dbc0b5433367b464c048NS-.
doi:10.1016/j.apacoust.2012.10.010, publisher: Elsevier
Ltd.
[39] M. Suhanek, S. GrubeÅąa, I. Djurek, A. PetoÅąic, in: Proceedings
of the 23rd International Congress on Acoustics, International
Commission for Acoustics (ICA), Aachen, Germany, 2019, pp.
884–890. doi:10.18154/RWTH-CONV- 239246.
[40] S. R. Payne, Applied Acoustics 74 (2013) 255–263. doi:10.1016/
j.apacoust.2011.11.005, publisher: Elsevier Ltd.
[41] S. R. Payne, C. Guastavino, Frontiers in Psychology 9 (2018) 1–17.
doi:10.3389/fpsyg.2018.02224.
[42] M. Lionello, F. Aletta, J. Kang, Applied Acoustics 170 (2020)
107479. URL: https://linkinghub.elsevier.com/retrieve/pii/
S0003682X20305831. doi:10.1016/j.apacoust.2020.
107479, publisher: Elsevier Ltd.
[43] Y. Hou, Q. Ren, H. Zhang, A. Mitchell, F. Aletta, J. Kang, D. Bot-
teldooren, The Journal of the Acoustical Society of America 154
(2023) 3145–3157. doi:10.1121/10.0022408, publisher: Acous-
tical Society of America.
[44] K. N. Watcharasupat, K. Ooi, B. Lam, T. Wong, Z.-T. Ong,
W.-S. Gan, IEEE Signal Processing Letters 29 (2022) 1749–
1753. URL: https://ieeexplore.ieee.org/document/9841611/.
doi:10.1109/LSP.2022.3194419.
[45] K. Ooi, K. N. Watcharasupat, B. Lam, Z.-T. Ong, W.-S. Gan, in:
ICASSP 2022 - 2022 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), IEEE, Singapore, 2022,
pp. 8887–8891. doi:10.1109/ICASSP43922.2022.9746897.
[46] K. Ooi, K. N. Watcharasupat, B. Lam, Z.-T. Ong, W.-S. Gan, in:
ICASSP 2023 - 2023 IEEE International Conference on Acous-
tics, Speech and Signal Processing (ICASSP), IEEE, Rhodes Is-
land, Greece, 2023, pp. 1–5. doi:10.1109/ICASSP49357.2023.
10094866.
[47] C. Moshona, F. Aletta, X. Chen, A. Fiebig, H. Henze, J. Kang,
A. Mitchell, T. Oberman, B. Schulte-Fortkamp, H. Tong,in: For um
Acusticum 2023, 10th Convention of the European Acoustics
Association, Turin, Italy. doi:10.61782/fa.2023.0087.
[48] K. Ooi, Z.-T. Ong, K. N. Watcharasupat, B. Lam, J. Y. Hong, W.-
S. Gan, IEEE Transactions on Affective Computing (2023) 1–17.
doi:10.1109/TAFFC.2023.3247914, arXiv: 2207.01078.
[49] T. Wong, K. N. Watcharasupat, B. Lam, K. Ooi, Z.-T. Ong,
F. A. Karnapi, W.-S. Gan, in: INTER-NOISE and NOISE-CON
Congress and Conference Proceedings, volume 265, Institute of
Noise Control Engineering, Glasgow, UK, 2022, pp. 2013–2021.
URL: http://arxiv.org/abs/2204.13890. doi:10.3397/IN_2022_
0290, arXiv: 2204.13890 Issue: 5 ISSN: 0736-2935.
[50] E. R. Thompson, Journal of Cross-Cultural Psychology 38 (2007)
227–242. URL: https://doi.org/10.1177/0022022106297301.
doi:10.1177/0022022106297301, publisher: SAGE
Publications Inc.
[51] N. D. Weinstein, Journal of Applied Psychology 63 (1978) 458–
466. URL: /record/1979-09992-001. doi:10.1037/0021-9010.
63.4.458.
[52] International Organization for Standardization, ISO/TS 15666
Acoustics âĂŤ Assessment of noise annoyance by means of social
and socio-acoustic surveys, International Organization for Stan-
dardization, Geneva, Switzerland, 2021.
[53] S. Cohen, T. Kamarck, R. Mermelstein, Journal of Health and
Social Behavior 24 (1983) 385. doi:10.2307/2136404.
[54] World Health Organization Regional Office for Europe, Wellbeing
measures in primary health care/the DepCare Project: report on a
WHO meeting, Technical Report WHO/EURO:1998-4234-43993-
62027, World Health Organization, Stockholm, Sweden, 1998.
URL: https://iris.who.int/handle/10665/349766.
[55] K. Ooi, Y. Xie, B. Lam, W.-S. Gan, MethodsX 8 (2021) 101288.
doi:10.1016/j.mex.2021.101288.
[56] W. Gao, J. Kang, H. Ma, C. Wang, Building and
Environment 245 (2023) 110945. URL: https://www.
sciencedirect.com/science/article/pii/S0360132323009721.
doi:10.1016/j.buildenv.2023.110945.
[57] C. Tarlao, J. Steffens, C. Guastavino, Building and Environment
188 (2021) 107490. doi:10.1016/j.buildenv.2020.107490,
publisher: Elsevier Ltd.
Lam et al.: Preprint submitted to Elsevier Page 39 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
[58] International Organization for Standardization, ISO 532-1:2017
- Acoustics âĂŤ Method for calculating loudness âĂŤ Part 1:
Zwicker method, volume 7, International Organization for Stan-
dardization, Geneva, Switzerland, 2017. Publication Title: ISO
532-1.
[59] D. Watson, L. A. Clark, A. Tellegen, Journal of Personality
and Social Psychology 54 (1988) 1063–1070. doi:10.1037/
0022-3514.54.6.1063.
[60] W. J. Conover, R. L. Iman, The American Statistician 35 (1981)
124. URL: https://www.jstor.org/stable/2683975?origin=crossref.
doi:10.2307/2683975.
[61] R Core Team, R: A Language and Environment for Statistical
Computing, 2023. URL: https://www.r-project.org/, place: Vienna,
Austria.
[62] A. Kuznetsova, P. B. Brockhoff, R. H. B. Christensen, Journal of
Statistical Software 82 (2017). doi:10.18637/jss.v082.i13.
[63] J. Fox, S. Weisberg, An {R} Companion to Applied Regression,
Sage, Thousand Oaks {CA}, 2019. URL: https://socialsciences.
mcmaster.ca/jfox/Books/Companion/.
[64] M. Ben-Shachar, D. LÃijdecke, D. Makowski, Journal of Open
Source Software 5 (2020) 2815. doi:10.21105/joss.02815.
[65] R. V. Lenth, emmeans: Estimated Marginal Means, aka Least-
Squares Means, 2023. URL: https://cran.r-project.org/package=
emmeans.
[66] G. Kidd, H. S. Colburn, in: J. C. Middlebrooks, J. Z. Simon, A. N.
Popper, R. R. Fay (Eds.), The Auditory System at the Cocktail
Party, Springer Cham, Cham, Switzerland, 2017, pp. 75–109.
doi:10.1007/978-3- 319-51662- 2_4.
[67] B. Jiang, W. Xu, W. Ji, G. Kim, M. Pryor, W. C. Sullivan, Journal
of Environmental Psychology 77 (2021) 101659. doi:10.1016/j.
jenvp.2021.101659, publisher: Academic Press.
[68] J. A. Benfield, P. A. Bell, L. J. Troup, N. C. Soderstrom, Journal of
Environmental Psychology 30 (2010) 103–111. doi:10.1016/j.
jenvp.2009.10.002, publisher: Elsevier Ltd.
[69] \. Axelsson, M. E. Nilsson, B. Berglund, The Journal of the
Acoustical Society of America 128 (2010) 2836–2846. URL:
http://asa.scitation.org/doi/10.1121/1.3493436. doi:10.1121/1.
3493436, publisher: Acoustical Society of America (ASA) ISBN:
9781441905604.
[70] J. Y. Jeon, H. I. Jo, K. Lee, Sustainable Cities and Society 99
(2023) 104929. URL: https://doi.org/10.1016/j.scs.2023.104929.
doi:10.1016/j.scs.2023.104929, publisher: Elsevier Ltd.
[71] J. Y. Hong, B. Lam, Z.-T. Ong, K. Ooi, W.-S. Gan, J. Kang,
S. Yeong, I. Lee, S.-T. T. Tan, Building and Environment 167
(2020) 106423. doi:10.1016/j.buildenv.2019.106423, pub-
lisher: Elsevier Ltd.
[72] M. Erfanian, A. Mitchell, F. Aletta, J. Kang, Journal of Environ-
mental Psychology 77 (2021) 101660. doi:10.1016/j.jenvp.
2021.101660, publisher: Elsevier Ltd.
[73] H. Li, S.-K. Lau, Applied Acoustics 166 (2020) 107372. doi:10.
1016/j.apacoust.2020.107372, publisher: Elsevier Ltd.
[74] F. A. Karnapi, B. Lam, T. Wong, K. Ooi, Z.-T. Ong, W.-S.
Gan, J. Hong, S. Yeong, in: INTER-NOISE and NOISE-CON
Congress and Conference Proceedings, volume 263, The Institute
of Noise Control Engineering of the USA, Inc., Washington, D.C.,
USA, 2021, pp. 2253–2258. doi:10.3397/IN-2021-2084, issue:
4 ISSN: 0736-2935.
[75] C. Moshona, F. Aletta, H. Henze, X. Chen, A. Mitchell, T. Ober-
man, H. Tong, A. Fiebig, J. Kang, B. Schulte-Fortkamp, in:
INTER-NOISE and NOISE-CON Congress and Conference Pro-
ceedings, volume 265, pp. 854–862. doi:10.3397/IN_2022_
0121, issue: 7 ISSN: 0736-2935.
[76] C. C. Moshona, S. Lepa, A. Fiebig, Applied Acoustics 207 (2023)
109338. doi:10.1016/j.apacoust.2023.109338.
[77] Y. Hou, S. Song, C. Luo, A. Mitchell, Q. Ren, W. Xie,
J. Kang, W. Wang, D. Botteldooren, in: INTERSPEECH 2023,
ISCA, ISCA, 2023, pp. 331–335. doi:10.21437/Interspeech.
2023-1021, arXiv: 2308.11980 ISSN: 19909772.
[78] A. Mitchell, T. Oberman, F. Aletta, M. Erfanian, M. Kachlicka,
M. Lionello, J. Kang, The International Soundscape Database: An
integrated multimedia database of urban soundscape surveys –
questionnaires with acoustical and contextual information (0.2.2)
[Data set], 2021. doi:10.5281/zenodo.5705908.
[79] A. Mitchell, F. Aletta, T. Oberman, M. Erfanian, J. Kang, INTER-
NOISE and NOISE-CON Congress and Conference Proceedings
268 (2023) 2108–2118. doi:10.3397/IN_2023_0309.
Lam et al.: Preprint submitted to Elsevier Page 40 of 41
Automating Urban Soundscape Enhancements with AI: In-situ Assessment of Quality and Restorativeness in
Traffic-Exposed Residential Areas
[80] L. Jiang, A. Bristow, J. Kang, F. Aletta, R. Thomas, H. Not-
ley, A. Thomas, J. Nellthorp, Building and Environment 219
(2022) 109231. doi:10.1016/j.buildenv.2022.109231, pub-
lisher: Elsevier Ltd.
[81] F. Aletta, J. Xiao, J. Kang, JASA Express Letters 4
(2024). URL: https://pubs.aip.org/jel/article/4/4/047401/
3280368/Identifying-barriers-to-engage-with-soundscape.
doi:10.1121/10.0025454.
Lam et al.: Preprint submitted to Elsevier Page 41 of 41
