Evaluation of Psychoacoustic Machine Learning Assessment Method for Predicting Occupant Well-Being

Abstract

In modern buildings, the air-conditioned indoor environment is vital for occupant productivity and well-being, yet fan noise and airflow turbulence can significantly compromise these benefits. Human–environmental interactions are complex processes that traditional energy-based acoustic metrics are often insufficient to model. Therefore, this study aims to advance the multidimensional sound quality assessment framework for building acoustics. Three methods, the conventional regression approach (CRA), general prediction model (GPM), and psychoacoustic machine learning (PML) assessment methods, were evaluated for predicting three perceptual dimensions (Evaluation, Potency, Activity; EPA) and negative noise impacts on occupant well-being (O1: Discomfortable, O2: Annoying, O3: Stressful, and O4: Unacceptable). Based on 432 multidimensional sound quality assessments across four general types of air-conditioned built environments, the PML achieved the best goodness-of-fit for the EPA-score perdition (adjusted R² = 0.61) compared to CRA (0.32) and GPM (0.15) and effectively predicted all negative noise impacts (adjusted R² = 0.53–0.61). The PML assessment method offers a smart and reliable solution for sound quality and well-being prediction through psychoacoustic heatmaps encoding time-varying psychoacoustic features in 227 × 227 pixels from 30 s soundtracks of the built environment for sustainable building design.

1. Introduction

The field of building acoustics [1,2] encompasses the assessment and control of noise within built environments, with particular emphasis on how environmental sound characteristics affect the well-being of building occupants [3]. Modern buildings increasingly rely on air-conditioning systems as integral components for maintaining satisfactory indoor environmental conditions [4]. These systems inevitably generate noise through multiple mechanisms, including mechanical fan operation and aerodynamic interactions where turbulent airflow encounters duct discontinuities [5,6]. Consequently, systematic acoustic assessment has emerged as a critical requirement for mitigating these issues [7]. Conventional acoustic assessment approaches [8,9] for air-conditioned built environments utilize established acoustic metrics including A-weighted equivalent continuous sound pressure level (L_Aeq), noise criteria (NC), noise rating (NR), and room criteria (RC). While these single-value metrics offer straightforward acceptability assessments of indoor noise conditions, they possess inherent limitations in characterizing the acoustic properties and complex psychological dimensions of occupant experience that extend beyond simple magnitude sensitivity [10]. Historically, noise assessment, prevention, prediction methodologies, and management frameworks have prioritized objective acoustic characterization, largely because objective metrics present fewer measurement and prediction challenges compared to subjective human responses [11]. Contemporary understanding emphasizes that human–environmental interactions constitute the foundation for advancing acoustic assessment in building acoustics. This recognition establishes multidimensional sound quality assessment [12] as fundamental for effectively analyzing human–environmental relationships and supporting evidence-based building design.

Public recognition of the health consequences [13] of daily noise exposure has increased following the update of the World Health Organization (WHO) noise guidelines in 2018 [14]. Daily noise exposure is linked to discomfort as well as stress biomarkers, disease progression, and mortality [15]. Undesired acoustic environments are recognized to cause stress [16], annoyance [17], and discomfort [18], all of which exert negative impacts on occupant well-being [19]. Human–environmental interactions with acoustic environments include sequential processes from sound generation and transmission to perception, ultimately concluding in diverse noise impact outcomes. Although ISO 15,666 [20] and ISO 12,913 standards [21,22,23] address subjective assessment of perceptual responses, these assessments exhibit substantial variability among building acoustics studies [24,25]. The semantic differential method [26], which employs questions constructed from bipolar semantic pairs representing opposing meanings [27,28], is a commonly adopted psychological scale for the quantitative evaluation of objective phenomena. A comprehensive systematic review [29] analyzing applications of the semantic differential method for subjective assessment discovered the fundamental structure of sound perception in terms of the Evaluation, Potency, and Activity (EPA) framework. The Evaluation (E-), Potency (P-), and Activity (A-) dimensions respectively represent general judgment, sensation of sound energy content, and sensation of the temporal and spectral content of sounds. Hence, acoustic metric selection for sound quality assessment should incorporate measurements addressing energy, spectral, and temporal characteristics of sound.

In objective acoustic characterization, psychoacoustic metrics comprise a collection of objective measures designed to quantify actual sound sensations, grounded in the psychoacoustic scale introduced in 1961 [30]. Total loudness (N) is the most widely recognized psychoacoustic metric, quantifying loudness perception with calculation procedures standardized in ISO 532-1 [31]. The N measurement provides valuable supplementation to conventional acoustic assessment methods for assessing indoor acoustic comfort [32] compared to NC and NR. In addition, sharpness (S), roughness (R), and fluctuation strength (FS) are usually accepted psychoacoustic metrics for evaluating the spectral properties of sounds. These metrics respectively quantify sharpness perception via energy skewness calculations, roughness perception through rapid amplitude modulation measurement, and fluctuation strength perception via slower amplitude modulation assessment [33]. Applying statistical sound level metrics such as L_A90 and L_A10, rather than time-equivalent levels, enables superior differentiation between ambient noise levels and event noise levels through consideration of temporal characteristics. These psychoacoustic metrics and statistical noise levels were also found to be significantly correlated with perceptual assessments across the E-, P-, and A-dimensions [34,35].

Contrasting with objective acoustic characterization, developing reliable, valid, and practical psychometric instruments for assessing multidimensional perceptual impacts remains an ongoing research priority. Various subjective assessment approaches, together with their theoretical factor models, have been proposed across different investigations [36]. However, analytical inconsistencies among these studies limit cross-study comparability. The Psychoacoustics Perception Scale (PPS) was therefore constructed based on the EPA framework to systematically evaluate all essential perceptions within the fundamental human perceptual dimensions of sound [12]. Identification of the EPA framework offers a promising approach for addressing challenges posed by extensive factor variability in multidimensional environmental sound quality assessment. Moreover, the EPA framework establishes a comparative structure across different studies and provides a basis for predicting negative noise impacts associated with occupant well-being. Recent research [37] has successfully applied multidimensional sound quality assessments to outdoor soundscapes by combining a comprehensive set of acoustic and psychoacoustic metrics with the PPS. This study evaluated the occupant well-being-related noise impacts (e.g., O1: Discomfortable, O2: Annoying, O3: Stressful, and O4: Unacceptable), which the WHO links to the disease burden of environmental noise. The results confirmed the value of the EPA framework in multidimensional sound quality assessments and facilitated the development of general prediction models (GPMs) that predictively model perceptual influences on occupant well-being.

Other than the conventional regression approach (CRA), the development of artificial neural networks (ANNs) with feature learning capabilities represents a significant advancement in machine learning (ML) algorithms for predictive modeling of perceptual influences on occupant well-being. While ANNs inspired by biological neural networks [38] have revolutionized fields like image classification [39], particularly through non-saturating ReLU activation functions [40], their application in psychoacoustics remains limited. A systematic search of the ScienceDirect database revealed that most “psychoacoustic machine learning” (PML) studies focused on objective signal classification tasks such as heartbeat analysis [41], vehicle diagnostics [42], and sound categorization [43,44,45], rather than on subjective human impact. Only a few studies have applied PML to predict subjective noise impacts [46,47,48], highlighting a gap in intelligent, holistic methodologies for assessing air-conditioned built environments.

To address the limitations of CRA assessment methods in capturing temporal variations and spectral complexities, a prior study [49] developed the psychoacoustic heatmap machine learning model (PHMLM). This innovative method incorporated the EPA framework to transform dynamic psychoacoustic metrics into detailed visual formats for forecasting noise effects. Unlike classical Mel-spectrograms [50], which display energy in constant frequency bands without accounting for human auditory sensations such as loudness and sharpness, the PHMLM utilized a preprocessing technique that reduced input data size by approximately 5000-fold. By encoding multidimensional acoustic information into 227 × 227-pixel psychoacoustic heatmaps with 0.002 s temporal resolution, the PHMLM efficiently captured energy, spectral, and temporal content for deep neural network analysis of multidimensional sound quality. This prior work [49] demonstrated that PML assessment methods could achieve predictive consistency with established regression methodologies (including GPM and CRA), providing a foundation for the comparative evaluation presented in this study.

The primary objective of this research is to evaluate different assessment methods (CRA, GPM, and PML) for predicting sound quality and subsequent negative noise impacts on occupant well-being in air-conditioned built environments using multidimensional objective acoustic characteristics. Detailed model development procedures have not been repeated in Section 2, as the historical progression of different methods used to analyze the complex interaction between the acoustic environment and occupant perception has been comprehensively presented in prior published studies (CRA [12,23], GPM [37], and PML [49]). Compared to these prior studies, this work delivers a comprehensive performance evaluation, cross-method comparison, and hybrid integration analysis through balanced sampling of air-conditioned environments, independent subjective perceptual response measurements beyond the original development datasets, and cross-domain generalization testing (outdoor GPM with indoor PML). These advances identify the relative strengths of each method and optimal integration strategies for building acoustic design. This significance of this research lies in delivering a smart, reliable, efficient, and user-friendly assessment method that can inform sustainable building design and enhance indoor acoustic environments by predicting occupant well-being responses to air-conditioning noise within built environments.

2. Materials and Methods

2.1. Multidimensional Sound Quality Assessment Framework

A holistic approach to sound quality assessment [12,37,49] is essential for capturing the complex interaction between the acoustic environment and occupant perception. As illustrated in Figure 1a, this study adopts a comprehensive framework comprising four key stages I-IV for objective and subjective data collection:

• Stage I. Objective acoustic characterization: This stage involves measuring the physical properties of sound using acoustic and psychoacoustic metrics (see Figure 1b) that capture energy content (e.g., L_Aeq, RC, N), spectral content (e.g., S, R, and FS), and temporal content (e.g., L_A10–L_A90 and N₅–N₉₅);
• Stage II. Fundamental perceptual responses: This stage measures the fundamental subjective perceptual responses to sound using the PPS, which consists of nine questions (E1: Quiet–Noisy, E2: Relaxed–Tense, E3: Pleasant–Unpleasant, P1: Quiet–Loud, P2: Light–Heavy, P3: Weak–Strong, A1: Deep–Metallic, A2: Low–High, and A3: Dull–Sharp) on a 7-level semantic differential scale (see Figure 1c);
• Stage III. EPA framework: This stage categorizes subjective responses into three fundamental dimensions (E, P, and A), serving as the core link in the human–environmental interaction process between physical sound stimuli and occupant perception (see Figure 1c);
• Stage IV. Occupant well-being-related noise impacts: The final stage evaluates the consequently negative impacts on occupant well-being (see Figure 1c), specifically Discomfort (O1), Annoying (O2), Stressful (O3), and Unacceptable (O4).

This multidimensional structure provides a standardized ground-truth for comparative analysis, ensuring that the CRA [12,23], GPM [37], and PML [49] assessment methods are evaluated against a consistent and holistic set of inputs from objective acoustic characterization to predict sound quality and subsequent negative noise impacts on occupant well-being (arrows in Figure 1a).

2.1.1. Objective Data Collection in Air-Conditioned Built Environments

To capture a representative sample of indoor acoustic conditions, sound recordings were conducted in four distinct air-conditioned built environments: (1) Library, (2) Classroom, (3) Lecture Hall, and (4) Office (see Figure 1b). These four air-conditioned built environments (Library, Classroom, Lecture Hall, and Office) represent essential daily spaces where sustained HVAC noise exposure critically impacts occupant concentration, learning, communication, and productivity. Libraries demand quiet for individual study and reading, classrooms support interactive teaching and group discussions, lecture halls accommodate large-group instruction and presentations, and offices facilitate collaborative work and focused tasks, each presenting distinct psychoacoustic challenges from HVAC systems. Their selection reflects prevalent built environments in educational institutions and workplaces where air-conditioning noise management directly influences occupant well-being and performance. These space types encompass the majority of indoor scenarios encountered daily by students, faculty, and professionals. Detailed acoustic characteristics of these environments were analyzed and compared in Section 3 to better understand the differences between environments. Soundtracks with an equal number from different air-conditioned built environments were randomly selected in this study from the prior PML development study [49]. Although identical to prior study recordings, independent listening tests were conducted to collect subjective perceptual responses, eliminating data leakage between the training phase of the prior study and the multidimensional sound quality assessments in this study. All 30 s soundtracks were recorded using a high-precision two-channel handheld analyzer (Type 2270; Brüel & Kjær, Nærum, Denmark) equipped with binaural recording microphones. The recording position was standardized at a height of 1.2 m above the floor to simulate the typical ear height of a seated occupant. A total of 432 30 s soundtracks were extracted from the recordings for both objective and subjective analysis of noise in air-conditioned built environments. Conventional acoustic metrics were recorded directly from the sound analyzer, while psychoacoustic metrics were computed using MATLAB R2024a (MathWorks, Natick, USA). The detailed calculation equations and unit definitions based on ISO 532-1 [31] and Zwicker’s book [30] are provided in Appendix A.

2.1.2. Subjective Perceptual Response Measurement

Jury listening tests were conducted to collect subjective responses to the recorded soundtracks (see Figure 1b). The tests were conducted in an anechoic chamber at the Hong Kong Polytechnic University to eliminate the influence of background noise and room reflections. To minimize the influence of confounding environmental factors, the anechoic chamber maintained a controlled temperature of 25 ± 2 °C with adequate lighting conditions. The 30 s soundtracks were replayed using an omni-directional dodecahedron loudspeaker (Type LS02; Acoem, Limonest, France) positioned 1.2 m from the listener, matching the original recording height. To ensure accurate reproduction, the sound pressure level (SPL) was calibrated using a 1 kHz reference tone at 60 dB measured at the listener’s ear position. Before the listening tests, written informed consent was obtained from all participants, and demographic details (gender and age) were documented. Individuals with self-reported hearing impairments or chronic noise exposure were excluded to ensure data validity. This criterion minimized potential biases from undetected auditory threshold shifts that could confound perceptual judgments.

Participants evaluated a total of 12 soundtracks using a self-administered questionnaire structured in two parts to capture the subjective responses defined in the assessment framework (see Figure 1c). All 12 soundtracks per participant were randomly assigned. Individual participant sets were not environment-balanced (due to randomization), but collectively ensured the same number of assessments per environment type. A 1 min idling period separated consecutive assessments. Part A employed the PPS to quantify sound quality perceptions across the three fundamental E-, P-, and A- dimensions. Part B assessed occupant well-being-related noise impact, where participants rated the specific negative outcomes (O1–O4) on a 7-point Likert scale. This dual-part instrument ensured that both the intermediate perceptual attributes and the consequential well-being impacts were systematically recorded for every acoustic stimulus.

2.2. Assessment Methods for Predictive Modeling of Subjective Responses

Three assessment methods (CRA, GPM, and PML) were applied and compared in this study to evaluate the feasibility of predicting occupant well-being.

2.2.1. Conventional Regression Approach (CRA)

The CRA represents the traditional methodology in environmental noise studies [23] involving subjective data. The objective acoustic and psychoacoustic metrics are directly used as independent variables in linear regression models (see Equation (1)) to predict subjective responses such as EPA-scores in the EPA framework.

$$y=B_1{x}_1+B_2{x}_2+B_3{x}_3+\cdots B_n{x}_n+C,$$

(1)

where y is the dependent variable, B₁–B_n are the regression coefficients (unstandardized coefficients) for the independent variables (x₁–x_n) as predictors, and C is the y-intercept.

2.2.2. General Prediction Model (GPM)

The fundamental distinction between the GPM and the CRA lies in their reliance on experimental data. While the CRA fits a regression model to the specific participant responses observed in this study, the GPM operates as a standalone predictive tool. In this context, the GPM is the “pure-objective” approach developed for occupant well-being prediction [37]. It derives predictions directly from acoustic and psychoacoustic metrics using established formulations, without further adjustment to the subjective data collected in this specific experiment.

By bypassing the need for dataset-specific correction, the GPM evaluates the generalizability and direct applicability of using established objective equations to predict subjective EPA-score (GPM_EPA, Equation (2)) and occupant well-being-related noise impacts (GPA_O1–4, Equations (3)–(6)). Although the GPM was originally formulated for broader applications in outdoor soundscape conditions, this study specifically applied the GPM to test its ability in indoor air-conditioned built environments.

$${\mathrm{G}\mathrm{P}\mathrm{M}}_{\mathrm{E}\mathrm{P}\mathrm{A}}=1.03\times \left(L_{\mathrm{A}10}-L_{\mathrm{A}90}\right)+0.68\times L_{\mathrm{A}50}-0.28\times N-30.4\times R-39.7,$$

(2)

$${\mathrm{GPM}}_{\mathrm{O}1: \mathrm{discomfortable} }=0.17\times {\mathrm{G}\mathrm{P}\mathrm{M}}_{\mathrm{E}\mathrm{P}\mathrm{A}}-5.51\times FS+3.7\times R_5+0.04\times L_{Zeq}-0.03\times L_{\mathrm{A}90}+2.06,$$

(3)

$${\mathrm{G}\mathrm{P}\mathrm{M}}_{\mathrm{O}2: \mathrm{A}\mathrm{n}\mathrm{n}\mathrm{o}\mathrm{y}\mathrm{i}\mathrm{n}\mathrm{g}}=0.18\times {\mathrm{G}\mathrm{P}\mathrm{M}}_{\mathrm{E}\mathrm{P}\mathrm{A}}-5.86\times FS-0.04\times f_{M\mathrm{o}\mathrm{d}}+4.77,$$

(4)

$${\mathrm{G}\mathrm{P}\mathrm{M}}_{\mathrm{O}3: \mathrm{S}\mathrm{t}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{f}\mathrm{u}\mathrm{l}}=0.18\times {\mathrm{G}\mathrm{P}\mathrm{M}}_{\mathrm{E}\mathrm{P}\mathrm{A}}+3.86,$$

(5)

$${\mathrm{G}\mathrm{P}\mathrm{M}}_{\mathrm{O}4: \mathrm{U}\mathrm{n}\mathrm{a}\mathrm{c}\mathrm{c}\mathrm{e}\mathrm{p}\mathrm{t}\mathrm{a}\mathrm{b}\mathrm{l}\mathrm{e}}=0.17\times {\mathrm{G}\mathrm{P}\mathrm{M}}_{\mathrm{E}\mathrm{P}\mathrm{A}}-5.58\times S+8.81\times R_5+3.77.$$

(6)

2.2.3. Psychoacoustic Machine Learning (PML) Assessment Method

The third and most advanced approach evaluated in this study is the PML assessment method, which utilizes the PHMLM [49], originally developed for EPA-score prediction. Unlike the equation-based GPM, the PML is a data-driven framework designed to capture complex, non-linear relationships between acoustic features and subjective perception.

For each 30 s soundtrack containing air-conditioning sounds, a corresponding psychoacoustic heatmap was generated as a 227 × 227-pixel grayscale image (see Figure 1b). These dimensions align with the input requirements of the AlexNet architecture [39] (see

Table 1. Architecture of a psychoacoustic heatmap machine learning model.

Layer	Configuration/Dimensions	Activation/Function
Input	227 × 227 × 1 (grayscale heatmap)	Input layer for psychoacoustic heatmaps
Convolution 1	96 filters, 11 × 11 kernel, stride 4	ReLU (Feature extraction—transferred)
Max Pooling 1	3 × 3 max pooling, stride 2	Down-sampling
Convolution 2	256 filters, 5 × 5 kernel, padding 2	ReLU (Feature extraction—transferred)
Max Pooling 2	3 × 3 max pooling, stride 2	Down-sampling
Convolution 3	384 filters, 3 × 3 kernel, padding 1	ReLU (Feature extraction—transferred)
Convolution 4	384 filters, 3 × 3 kernel, padding 1	ReLU (Feature extraction—transferred)
Convolution 5	256 filters, 3 × 3 kernel, padding 1	ReLU (Feature extraction—transferred)
Max Pooling 3	3 × 3 max pooling, stride 2	Down-sampling
Fully Connected 6	4096 neurons	ReLU, Dropout (0.5) (High-level feature processing)
Fully Connected 7	4096 neurons	ReLU, Dropout (0.5) (High-level feature processing)
Fully Connected 8	19 neurons	Softmax (classification for −9 to +9 scale)
Output	19-level probability distribution	E-/P-/A-score prediction

Notes. ReLU = Rectified Linear Unit activation function; Dropout = regularization technique to prevent overfitting; E = Evaluation; P = Potency; A = Activity. The final predicted EPA-score is derived by combining the separate outputs of three parallel models independently trained for E-, P, and A-scores.

). Detailed psychoacoustic heatmap generation procedures, including matrix selection and resolution rationale, can be found in the study of PML development [49]. The heatmap encodes temporal variations in the four most commonly applied psychoacoustic metrics in environmental noise studies (N, R, S, and FS), calculated with a time resolution of 0.002 s. This resolution utilizes 96 samples at a 48 kHz sample rate, following ISO 532-1 standards. The specific covering durations and maximum intensities mapped to the pixel value of 255 (white) were set as follows:

• Rows 1–57: N covering 25.878 s, with a maximum intensity of 25.5 sone (0.1 sone/pixel);
• Rows 58–114: S covering 25.878 s, with a maximum intensity of 2.55 acum (0.01 acum/pixel);
• Rows 115–171: R covering 25.878 s, with a maximum intensity of 0.31875 asper (0.00125 asper/pixel);
• Rows 172–227: FS covering 25.424 s, with a maximum intensity of 0.255 vacil (0.001 vacil/pixel).

For this comparative study, the PML model was deployed using the pre-trained PHMLM-EPA established in prior work [49]. As the training phase (including hyperparameter optimization and stochastic gradient descent with momentum (SGDM) solver configuration) was completed and validated in the prior study, this research utilizes the fully trained network directly to generate predictions for the current dataset. This ensures that the comparison between PML, GPM, and CRA focuses strictly on predictive performance.

2.3. Statistical Analysis

All statistical analysis were executed using IBM SPSS Statistics 26 (IBM Corp., Armonk, NY, USA) at a significance level of 0.05 [51]. As a preliminary step, the Shapiro–Wilk test [52] was applied to assess the normality of the dataset. Given that the acoustic metrics and subjective responses demonstrated non-normal distributions (p-values < 0.05), non-parametric statistical methods were adopted for the two-tailed statistical tests in this study. To characterize the acoustic variations within the study, the Kruskal–Wallis test was employed to evaluate statistically significant differences in acoustic characteristics and subjective responses across the different types of air-conditioned built environments. Upon the detection of significant variances (p-values < 0.05), Dunn’s post hoc test with Bonferroni adjustments was performed to identify specific pairwise differences between different built environments. The internal consistency of the semantic differential scales was validated using Cronbach’s α (> 0.60) [53], justifying the reliability of combining individual items into composite E-, P-, and A-scores.

The comparative evaluation of the predictive modeling approaches (CRA, GPM, and PML) focused on two key performance indicators: Spearman’s rank correlation coefficient (ρ) and the adjusted coefficient of determination (adjusted R²). Spearman’s ρ was applied to test the strength and direction of the monotonic relationships (from −1 to 1) between the model-predicted values and the ground-truth experimental data, both for EPA-score and occupant well-being-related noise impacts (O1–O4).

$$\rho =1-{\displaystyle \frac{6\sum d_i^2}{n(n^2-1)}},$$

(7)

where n is the total sample size and d_i is the difference between the two ranks of each sample.

The adjusted R² served as a metric for goodness-of-fit, quantifying the proportion of variance in the subjective responses that could be explained by each model while penalizing model complexity. A higher adjusted R² value indicates superior explanatory power, allowing comparison between the simpler regression-based approaches (CRA and GPM) and the more complex PML assessment methods.

$$Adjusted R^2=1-\left(1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(\widehat{y_i}-\overline{y}\right)}^2 }{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2 }}\right)\times {\displaystyle \frac{(n-1)}{(n-p-1)}},$$

(8)

where n is the total sample size, p is the number of predictors, and $\widehat{y_i}$ and $\overline{y}$ are the predicted values and the mean of y, respectively.

3. Results

3.1. Descriptive Statistics of Multidimensional Sound Quality Assessment

A total of 432 jury listening tests were conducted for analysis, derived from 36 participants (69.4% male, 30.6% female; mean age 24.1 ± 2.2 years) evaluating 12 soundtracks each. The number of sound quality assessment in this study was balanced across four different types of air-conditioned built environments: (1) Library, (2) Classroom, (3) Lecture Hall, and (4) Office, with 108 assessments per type.

3.1.1. Multidimensional Objective Acoustic Characterization

The descriptive statistics for the objective acoustic and psychoacoustic metrics across the four environment types are summarized in

Table 2. Descriptive statistics of the acoustic characteristics of the soundtracks (n = 432) recorded from air-conditioned built environments in terms of nine conventional acoustic metrics.

Metric	Unit	1. Library (n = 108)	2. Classroom (n = 108)	3. Lecture Hall (n = 108)	4. Office (n = 108)	Total (n = 432)	Kruskal–Wallis Test	Post Hoc Tests
		Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	p-Value	Order
LZeq	dB	66.1 (8.6)	72.4 (5.8)	72.7 (8.2)	64.4 (9.5)	68.4 (8.3)	<0.001	4 < 1 < 2 = 3
LAeq	dBA	45.1 (7.5)	48.6 (8.6)	49.4 (8.7)	50.0 (7.0)	48.4 (8.6)	<0.001	1 < 2 = 3 = 4
LA10	dBA	45.5 (7.5)	49.2 (9.2)	50.0 (8.1)	50.3 (7.3)	48.9 (8.6)	<0.001	1 < 2 = 3 = 4
LA50	dBA	45.1 (7.5)	48.5 (8.5)	49.3 (9.0)	50.0 (7.0)	48.4 (8.6)	<0.001	1 < 2 = 3 = 4
LA90	dBA	44.7 (7.4)	47.9 (8.1)	48.9 (9.3)	49.6 (7.5)	48.0 (8.3)	<0.001	1 < 2 = 3 = 4
LA10–LA90	dBA	0.75 (0.2)	0.81 (0.4)	1.1 (0.6)	0.70 (0.2)	0.79 (0.4)	<0.001	1 = 4 < 2 = 3
NC	NC	43 (9)	46 (9)	47 (8)	46 (8)	46 (8)	<0.001	1 < 2 < 3 = 4
NR	NR	43 (8)	46 (9)	48 (7)	47 (7)	46 (9)	<0.001	1 < 2 < 3 = 4
RC	RC	42 (8)	45 (8)	47 (8)	47 (8)	45 (8)	<0.001	1 < 2 = 3 = 4

Notes. n = number of soundtracks, L_Z = unweighted sound pressure level, L_A = A-weighted sound pressure level, NC = noise criterion, NR = noise rating, RC = room criteria, _eq = time-equivalent, _{5/10/50/90/95} = percentiles of 5/10/50/90/95%, IQR = interquartile range, post hoc tests = Dunn’s tests of an acoustic metric in multiple comparisons between different environmental types.

and

Table 3. Descriptive statistics of the acoustic characteristics of the soundtracks (n = 432) recorded from air-conditioned built environments in terms of fifteen psychoacoustic metrics.

Metric	Unit	1. Library (n = 108)	2. Classroom (n = 108)	3. Lecture Hall (n = 108)	4. Office (n = 108)	Total (n = 432)	Kruskal–Wallis Test	Post Hoc Tests
		Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	Median (IQR)	p-Value	Order
LN	phon	64.5 (8.7)	67.4 (8.7)	69.1 (9.6)	83.1 (8.1)	70.2 (13.5)	<0.001	1 < 2 < 3 < 4
N	sone	5.5 (3.5)	6.7 (3.8)	7.5 (5.0)	19.9 (10.8)	8.1 (7.6)	<0.001	1 < 2 < 3 < 4
N5	sone	5.7 (3.6)	6.9 (4.0)	7.9 (4.8)	20.7 (11.0)	8.4 (7.6)	<0.001	1 < 2 < 3 < 4
N95	sone	5.1 (3.3)	6.2 (3.6)	6.7 (4.9)	18.1 (9.9)	7.5 (7.0)	<0.001	1 < 2 = 3 < 4
S	acum	1.2 (0.1)	0.87 (0.2)	1.1 (0.2)	1.2 (0.1)	1.1 (0.3)	<0.001	2 < 3 < 1 < 4
S5	acum	1.3 (0.1)	0.95 (0.3)	1.2 (0.3)	1.3 (0.1)	1.2 (0.3)	<0.001	2 < 3 < 1 < 4
S95	acum	1.1 (0.1)	0.81 (0.2)	0.90 (0.3)	1.2 (0.1)	1.0 (0.3)	<0.001	2 < 3 < 1 < 4
R	asper	0.07 (0.02)	0.07 (0.03)	0.08 (0.02)	0.12 (0.05)	0.08 (0.03)	<0.001	1 = 2 = 3 < 4
R5	asper	0.11 (0.03)	0.11 (0.04)	0.12 (0.03)	0.19 (0.10)	0.12 (0.05)	<0.001	1 = 2 < 3 < 4
R95	asper	0.04 (0.01)	0.04 (0.02)	0.05 (0.01)	0.07 (0.03)	0.05 (0.02)	<0.001	1 = 2 = 3 < 4
FS	vacil	0.01 (0.04)	0.03 (0.02)	0.02 (0.03)	0.06 (0.05)	0.03 (0.04)	<0.001	1 = 3 < 2 < 4
FS5	vacil	0.02 (0.05)	0.04 (0.03)	0.03 (0.03)	0.08 (0.06)	0.04 (0.05)	<0.001	1 < 3 < 2 < 4
FS95	vacil	0.005 (0.03)	0.02 (0.02)	0.007 (0.02)	0.05 (0.04)	0.02 (0.03)	<0.001	2 < 1 = 3 < 4
fMod	Hertz	135 (48)	133 (55)	141 (35)	89 (70)	131 (48)	<0.001	4 < 1 = 2 = 3
N5–N95	sone	0.64 (0.3)	0.78 (0.4)	1.1 (0.4)	2.0 (1.2)	0.96 (0.7)	<0.001	1 < 2 < 3 < 4

Notes. n = number of soundtracks, _{5/10/50/90/95} = percentiles of 5/10/50/90/95%, IQR = interquartile range, N = total loudness, S = sharpness, R = roughness, FS = fluctuation strength, f_Mod = dominant modulation frequency determined in the calculation of R, post hoc tests = Dunn’s tests of an acoustic metric in multiple comparisons between different environmental types.

. Shapiro–Wilk normality tests indicated that none of the acoustic or psychoacoustic metrics followed a normal distribution (p-values < 0.05), requiring the use of non-parametric statistical analysis. The results of the Kruskal–Wallis tests showed significant differences (p-values < 0.001) in acoustic energy, spectral, and temporal content among the environment types.

The built environment of l. Library consistently exhibited the lowest noise levels across most conventional metrics, including L_Aeq (Median = 45.1 dBA), L_A10 (45.5 dBA), and RC ratings (42). Post hoc Dunn’s tests confirmed that the Library environment was significantly quieter than the other three types. The built environments of 2. Classroom and 3. Lecture Hall demonstrated significantly higher temporal variability (L_A10–L_A90 = 1.1 dBA) compared to the other environment types. The results suggested that although the built environment of 4. Office is louder than the Library, the noise profile is more constant than that of the university Classroom and Lecture Hall.

Similar to the conventional acoustic metrics, the psychoacoustic characteristics exhibited significant variations across the four environment types (p-values < 0.001). The built environment of 4. Office was characterized by the highest intensities across most psychoacoustic metrics. It recorded the highest median N (19.9 sone), R (0.12 asper), and FS (0.06 vacil). Post hoc Dunn’s tests confirmed that the Office environment was significantly louder and rougher than all other types. In addition, the Office environment illustrated the highest temporal variability in loudness fluctuation (N₅–N₉₅ = 2.0 sone). In contrast, the built environment of 2. Classroom, despite being relatively quiet (N = 6.7 sone), was distinguished by its spectral characteristics. It exhibited the lowest S (0.87 acum), suggesting a duller sound profile compared to other built environments.

3.1.2. Multidimensional Subjective Perceptual Responses

The subjective assessments quantified via the PPS in the EPA framework are presented in Figure 2. Significant differences were observed for all perceptual dimensions across the four environment types (p-values < 0.05). The built environment of 1. Library was consistently rated as the most favorable environment, achieving the lowest EPA-score (−7.0 ± 4.8). In the design of the PPS, lower values indicate better sound quality (e.g., −3 = “Extremely Pleasant” vs. +3 = “Extremely Unpleasant”). Therefore, the Library environment was perceived as the “Quietest” (E1/P1), most “Relaxed” (E2), and most “Pleasant” (E3). Moreover, the built environment of 2. Classroom showed the lowest A-score (−2.6 ± 4.0), a result that aligns with its lowest objective S value (0.87 acum) in Table 2.

3.1.3. Reliability of Sound Quality Assessment

Cronbach’s α analysis was conducted to verify the internal consistency of the subjective responses measured by the PPS in quantifying sound quality perceptions across the three fundamental E-, P-, and A-dimensions. As shown in

Table 4. Cronbach’s α coefficients for subjective responses across four types of air-conditioned built environments.

Environment Type	Evaluation (E)	Potency (P)	Activity (A)	EPA
1. Library (n = 108)	0.969	0.961	0.943	0.977
2. Classroom (n = 108)	0.929	0.877	0.945	0.916
3. Lecture Hall (n = 108)	0.915	0.875	0.880	0.872
4. Office (n = 108)	0.888	0.829	0.863	0.827
Total (n = 432)	0.934	0.897	0.918	0.923

, the Cronbach’s α for all perceptual E-, P-, A-, and EPA-dimensions exceeded 0.80 across all built environments, indicating good internal consistency in the collected data [53].

3.2. Performance Evaluation of Assessment Methods

3.2.1. Prediction of Individual E-, P-, and A-Scores Using the PML Assessment Method

As shown in

Table 5. Spearman’s rank correlation coefficients between the individual E-, P-, and A-scores and the psychoacoustic machine learning (PML) predictions and the individual acoustic and psychoacoustic metrics.

Variable	PML	LAeq	LA10	LA50	LA90	NC	NR	RC	LN	N	N5	N95	S	S5	S95	R	R5	R95	fMod	N5–N95
E-score	0.77	0.31	0.56	0.56	0.56	0.56	0.55	0.56	0.56	0.38	0.38	0.38	NA	NA	NA	0.31	0.30	0.30	NA	0.33
P-score	0.80	0.30	0.58	0.58	0.58	0.58	0.58	0.58	0.58	0.37	0.37	0.37	NA	NA	NA	0.26	0.25	0.26	NA	0.32
A-score	0.57	NA	0.16	0.16	0.16	0.16	0.12	0.14	0.16	0.21	0.21	0.21	0.21	0.17	0.14	0.17	0.18	0.18	0.17	0.17

Notes. n = number of soundtracks, L_Z = unweighted sound pressure level, L_A = A-weighted sound pressure level, NC = noise criterion, NR = noise rating, RC = room criteria, _eq = time-equivalent, _{5/10/50/90/95} = percentiles of 5/10/50/90/95%, N = total loudness, S = sharpness, R = roughness, FS = fluctuation strength, f_Mod = dominant modulation frequency determined in the calculation of R. NA = no significant correlation found.

, higher ρ values were found between the PML predictions and the individual E-, P-, and A-scores compared to conventional metrics. For the E-score, the PML prediction (ρ = 0.77) showed a 148% and 38% improvement over L_Aeq (ρ = 0.31) and RC (ρ = 0.56), respectively. Similarly, for the P-score, the PML prediction (ρ = 0.80) demonstrated a 166% and 38% improvement over L_Aeq (ρ = 0.30) and RC (ρ = 0.58). For the A-score, the PHMLM (ρ = 0.57) showed a 171% improvement over the psychoacoustic metric S (ρ = 0.21).

3.2.2. Prediction of EPA-Score

The CRA_EPA model was established using stepwise linear regression to identify the most significant acoustic and psychoacoustic metrics as predictors for the EPA-score. The analysis identified RC, N₅−N₉₅, f_Mod, L_Zeq and L_A90 as the significant independent variables. The final regression model achieved an adjusted R² = 0.315, F(5, 426) = 40.6, p-value < 0.001). Among the predictors, RC demonstrated the strongest positive influence (B = 3.96, p-value < 0.001), while N₅−N₉₅ showed a strong negative contribution (B = −4.16, p-value < 0.001). The background noise level (L_A90) also negatively impacted the EPA-score (B = −2.19, p-value = 0.006), with minor contributions from L_Zeq (B = −0.40, p-value = 0.001) and f_Mod (B = −0.03, p-value = 0.013). The resulting regression equation was found to be:

$${\mathrm{C}\mathrm{R}\mathrm{A}}_{\mathrm{E}\mathrm{P}\mathrm{A}}=3.96\times RC-4.16\times \left(N_5-N_{95}\right)-0.03\times f_{\mathrm{M}\mathrm{o}\mathrm{d}}-0.40\times L_{\mathrm{Z}\mathrm{e}\mathrm{q}}-2.9\times L_{\mathrm{A}90}-41.7$$

(9)

The predictive performance of the overall sound quality (EPA-score) was evaluated by comparing the PML assessment methods against the CRA_EPA (Equation (9)), GPM_EPA (Equation (2)), and the metrics L_Aeq, RC, and N. As shown in Figure 3, the PML_EPA demonstrated the highest goodness-of-fit among the models across all environment types. For all air-conditioned built environments, the PML_EPA achieved a strong correlation (ρ = 0.79) and explained the highest variance (adjusted R² = 0.61). In contrast, the CRA_EPA showed moderate performance (ρ = 0.58, adjusted R² = 0.32), while the GPM_EPA showed the weakest predictive power (ρ = 0.40, adjusted R² = 0.15). The performance of the GPM_EPA was comparable to that of single-value acoustic metrics such as N (ρ = 0.40), indicating that general models fail to capture the specific acoustic properties in different environments.

The superiority of the PML approach was consistent across individual environments. In the built environment of 3. Lecture Hall, the PML_EPA achieved an exceptional fit (adjusted R² = 0.77), providing a 28% improvement over the CRA_EPA (adjusted R² = 0.60) and a 43% improvement over the GPM_EPA (adjusted R² = 0.54). Similarly, in the built environment of 1. Library, the PML_EPA (adjusted R² = 0.69) outperformed the CRA_EPA (adjusted R² = 0.23) by 200%, highlighting the strength of the heatmap-based features in quiet environments where traditional metrics often struggle. Furthermore, traditional acoustic metrics (L_Aeq, RC, and N) showed limited predictive capability, with adjusted R² values generally below 0.30 for all air-conditioned environment (see Figure 3d–f). The results confirmed that while conventional metrics and CRA_EPA can provide a baseline assessment, they lack the sensitivity of the PML_EPA to accurately predict multidimensional sound quality in complex air-conditioned environments.

3.2.3. Prediction of Occupant Well-Being-Related Noise Impacts

The capability of the models to predict occupant well-being, specifically the negative noise impacts (O1: Discomfortable, O2: Annoying, O3: Stressful, and O4: Unacceptable), was evaluated by integrating the predicted EPA-scores into the GPM_O1–4, Equations (3)–(6). Two hybrid models were established: GPM_O1–4-PML, which substitutes the GPM_EPA in the equations with the PML_EPA, and GPM_O1–4-PCRA, which substitutes it with the CRA_EPA. These hybrid models were then compared against the standalone PML_EPA and CRA_EPA, as well as traditional acoustic metrics (L_Aeq, RC, and N), to assess their predictive performance for occupant well-being-related noise impacts.

Table 6. Comparison of Spearman’s rank correlation coefficients ρ and adjusted coefficients of determination (adjusted R² in parentheses) for predicting occupant well-being related noise impacts (O1: Discomfortable, O2: Annoying, O3: Stressful, and O4: Unacceptable) in the four types of air-conditioned built environment using different assessment methods.

Dependent Variable	PMLEPA	CRAEPA	GPMO1–4	GPMO1–4-PML	GPMO1–4-CRA	LAeq	RC	N
O1: Discomfortable
1. Library (n = 108)	0.74 (0.53)	0.45 (0.18)	0.32 (0.10)	0.73 (0.53)	0.46 (0.20)	0.35 (0.11)	0.44 (0.21)	0.43 (0.16)
2. Classroom (n = 108)	0.49 (0.23)	0.38 (0.14)	0.31 (0.09)	0.48 (0.22)	0.39 (0.13)	0.42 (0.15)	0.66 (0.45)	0.65 (0.44)
3. Lecture Hall (n = 108)	0.66 (0.49)	0.71 (0.50)	0.57 (0.26)	0.65 (0.49)	0.71 (0.49)	0.64 (0.42)	0.40 (0.14)	0.41 (0.16)
4. Office (n = 108)	0.66 (0.43)	0.50 (0.26)	0.36 (0.11)	0.66 (0.43)	0.46 (0.23)	0.43 (0.20)	0.37 (0.14)	0.35 (0.10)
Total (n = 432)	0.65 (0.42)	0.50 (0.25)	0.32 (0.09)	0.64 (0.41)	0.49 (0.24)	0.48 (0.22)	0.48 (0.23)	0.33 (0.06)
O2: Annoying
1. Library (n = 108)	0.78 (0.59)	0.45 (0.20)	0.42 (0.13)	0.70 (0.49)	0.44 (0.21)	0.38 (0.12)	0.39 (0.15)	0.36 (0.11)
2. Classroom (n = 108)	0.46 (0.19)	0.35 (0.12)	0.30 (0.08)	0.27 (0.08)	0.23 (0.05)	0.38 (0.12)	0.36 (0.11)	0.37 (0.12)
3. Lecture Hall (n = 108)	0.67 (0.47)	0.71 (0.48)	0.56 (0.17)	0.58 (0.31)	0.54 (0.22)	0.67 (0.43)	0.68 (0.44)	0.67 (0.45)
4. Office (n = 108)	0.61 (0.36)	0.55 (0.30)	0.22 (0.05)	NA	NA	0.48 (0.23)	0.48 (0.24)	0.48 (0.21)
Total (n = 432)	0.63 (0.40)	0.50 (0.25)	0.42 (0.16)	0.35 (0.13)	0.20 (0.04)	0.46 (0.20)	0.46 (0.21)	0.18 (0.01)
O3: Stressful
1. Library (n = 108)	0.76 (0.62)	0.45 (0.20)	0.35 (0.15)	0.76 (0.62)	0.45 (0.20)	0.37 (0.12)	0.37 (0.15)	0.35 (0.11)
2. Classroom (n = 108)	0.60 (0.33)	0.46 (0.20)	0.25 (0.03)	0.60 (0.33)	0.46 (0.20)	0.46 (0.19)	0.44 (0.18)	0.46 (0.19)
3. Lecture Hall (n = 108)	0.50 (0.27)	0.46 (0.24)	0.15 (0.02)	0.50 (0.27)	0.46 (0.24)	0.44 (0.21)	0.45 (0.22)	0.45 (0.24)
4. Office (n = 108)	0.61 (0.38)	0.50 (0.26)	−0.27 (0.08)	0.61 (0.38)	0.50 (0.26)	0.45 (0.21)	0.46 (0.22)	0.46 (0.20)
Total (n = 432)	0.64 (0.41)	0.46 (0.22)	0.17 (0.03)	0.64 (0.41)	0.46 (0.22)	0.42 (0.17)	0.43 (0.18)	0.23 (0.02)
O4: Unacceptable
1. Library (n = 108)	0.77 (0.61)	0.40 (0.15)	0.29 (0.08)	0.75 (0.57)	0.46 (0.22)	0.32 (0.09)	0.33 (0.11)	0.31 (0.07)
2. Classroom (n = 108)	0.50 (0.24)	0.43 (0.19)	0.34 (0.12)	0.39 (0.15)	0.27 (0.06)	0.45 (0.19)	0.44 (0.18)	0.44 (0.19)
3. Lecture Hall (n = 108)	0.69 (0.48)	0.65 (0.45)	0.65 (0.23)	0.56 (0.33)	0.49 (0.25)	0.66 (0.45)	0.68 (0.46)	0.66 (0.47)
4. Office (n = 108)	0.39 (0.15)	0.32 (0.10)	0.22 (0.05)	0.36 (0.13)	0.32 (0.10)	0.26 (0.07)	0.27 (0.08)	0.27 (0.07)
Total (n = 432)	0.60 (0.36)	0.45 (0.20)	0.42 (0.15)	0.50 (0.25)	0.36 (0.13)	0.44 (0.19)	0.44 (0.19)	0.37 (0.12)

Notes. n = number of soundtracks, CRA_EPA = conventional regression approach for EPA-score prediction, PML_EPA = psychoacoustic machine learning for EPA-score prediction, GPM_O1–4 = general prediction model for O1–O4 prediction, GPM_O1–4-PML = hybrid models combining GPM_O1–4 and PML_EPA, GPM_O1–4-CRA = hybrid models combining GPM_O1–4 and CRA_EPA, L_Aeq = A-weighted equivalent continuous sound pressure level, RC = room criteria, N = total loudness, NA = no significant correlation found.

and

Table 7. Comparison of root mean squared error (RMSE) and confidence intervals (CIs) of predicted occupant well-being-related noise impacts (O1: Discomfortable, O2: Annoying, O3: Stressful, and O4: Unacceptable) in the four types of air-conditioned built environments using different assessment methods.

Dependent Variable	PMLEPA	CRAEPA	GPMO1–4	GPMO1–4-PML	GPMO1–4-CRA	LAeq	RC	N
O1: Discomfortable (n = 432, Mean = 3.87, SD = 1.52)
RMSE	1.16	1.32	1.45	1.17	1.33	1.34	1.33	1.47
95% CIs	[3.77, 3.99]	[3.75, 4.00]	[3.74, 4.01]	[3.76, 3.99]	[3.75, 4.00]	[3.75, 4.00]	[3.75, 4.00]	[3.74, 4.01]
O2: Annoying (n = 432, Mean = 3.91, SD = 1.49)
RMSE	1.15	1.29	1.36	1.39	1.46	1.33	1.32	1.48
95% CIs	[3.8, 4.02]	[3.79, 4.03]	[3.78, 4.04]	[3.78, 4.04]	[3.77, 4.05]	[3.79, 4.04]	[3.79, 4.04]	[3.77, 4.05]
O3: Stressful (n = 432, Mean = 3.70, SD = 1.43)
RMSE	1.10	1.27	1.43	1.10	1.27	1.30	1.29	1.42
95% CIs	[3.59, 3.8]	[3.58, 3.82]	[3.56, 3.83]	[3.59, 3.08]	[3.58, 3.82]	[3.57, 3.82]	[3.57, 3.82]	[3.56, 3.83]
O4: Unacceptable (n = 432, Mean = 3.56, SD = 1.53)
RMSE	1.17	1.32	1.41	1.26	1.38	1.36	1.35	1.51
95% CIs	[3.44, 3.67]	[3.43, 3.69]	[3.42, 3.70]	[3.44, 3.68]	[3.42, 3.69]	[3.43, 3.69]	[3.43, 3.69]	[3.41, 3.71]

Notes. n = number of soundtracks, CRA_EPA = conventional regression approach for EPA-score prediction, PML_EPA = psychoacoustic machine learning for EPA-score prediction, GPM_O1–4 = general prediction model for O1–O4 prediction, GPM_O1–4-PML = hybrid models combining GPM_O1–4 and PML_EPA, GPM_O1–4-CRA = hybrid models combining GPM_O1–4 and CRA_EPA, L_Aeq = A-weighted equivalent continuous sound pressure level, RC = room criteria, N = total loudness.

demonstrated a consistent relationship where higher adjusted R² corresponds to lower root mean squared error (RMSE) and narrower 95% confidence intervals (CIs), confirming the reliability of adjusted R² as the primary statistics indicator across all methods for subsequent discussion in this paper. As shown in Table 6, the PML_EPA consistently achieved the highest goodness-of-fit across all well-being-related noise impacts (O1–O4) for all air-conditioned built environments (n = 432). For O1: Discomfortable, the PML_EPA achieved a correlation of ρ = 0.65 (adjusted R² = 0.42), significantly outperforming the CRA_EPA (ρ = 0.50, adjusted R² = 0.25). The hybrid model GPM_O1-PML yielded nearly identical performance to the standalone PML_EPA (ρ = 0.64, adjusted R² = 0.41), whereas the GPM_O1-CRA and traditional acoustic metrics (L_Aeq and RC) showed weaker correlations (ρ = 0.48). The superiority of the PML assessment method was even more pronounced for O2: Annoying and O3: Stressful. The PML_EPA achieved correlations of 0.63 and 0.64 (adjusted R² = 0.40–0.41), respectively, providing a substantial improvement over the CRA_EPA (ρ = 0.46–0.50). In contrast, the GPM_O2–O3 and single-value metrics (L_Aeq, RC, and N) showed weak predictive power, with adjusted R² values dropping as low as 0.02 for O3: Stressful predictions using N. For O4: Unacceptable, the PML_EPA remained the best predictor (ρ = 0.60, adjusted R² = 0.36), effectively identifying the acoustic conditions that occupants deemed unacceptable. Notably, in the library environment, the PML_EPA achieved exceptional accuracy across all dimensions, with correlations reaching ρ = 0.74–0.78 (adjusted R = 0.36 = 0.53–0.62), demonstrating its sensitivity to the subtle acoustic properties that drive dissatisfaction in quiet spaces. These results confirm that integrating the multidimensional EPA-scores predicted by the PML into well-being models GPM_O1–4-PML offers a far more reliable tool for acoustic assessment than CRA or simple energy-based acoustic or psychoacoustic metrics.

4. Discussion

4.1. Multidimensional Sound Quality Assessment Framework Application

The fundamental advantage of the PML assessment method lies in its integration of “psychoacoustics” with “machine learning” to create a holistic, multidimensional sound quality assessment framework (see Figure 1). The results of the multidimensional objective acoustic characterization demonstrated that psychoacoustic metrics (see Table 3) provide critical spectral and temporal information that complements traditional energy-based metrics (see Table 2). For instance, while traditional metrics indicated varying noise levels, the psychoacoustic analysis revealed that the built environment of 2. Classroom exhibited the lowest S values. This spectral characteristic, which traditional metrics failed to capture, aligns perfectly with the lower A-scores observed in the subjective responses (see Figure 2), confirming the agreement between objective psychoacoustic features and human perception. This supports recent findings that single-number ratings often fail to capture the acoustic characteristics of noise perception in complex indoor environments [54].

Furthermore, the study demonstrated that subjective perceptual responses cannot be explained solely by energy content-related metrics. A striking example is the built environment of 4. Office, which recorded the highest energy values in Table 2. Counter-intuitively, the subjective results in Figure 2 showed that the office environment had a P-score of −1, which is lower than that of the 2. Classroom and 3. Lecture Hall. This discrepancy can be attributed to the specific psychoacoustic characteristics of the Office noise, specifically the lowest f_Mod observed in Table 3. The low loudness sensation associated with low-frequency noise leads to a lower perception of Potency despite high physical energy levels, a phenomenon consistent with studies on low-frequency HVAC noise impacts [55]. Consequently, this resulted in a lower overall EPA-score for the office environment compared to university Classroom and Lecture Hall environments. These findings substantiate the necessity of using the PPS for multidimensional sound quality assessment, as simple energy metrics fail to account for the complex non-linear processing of human hearing [56]. The high internal consistency (Cronbach’s α > 0.82) observed across all environment types further validates the PPS as a reliable psychometric tool for evaluating air-conditioned built environments [12,49].

4.2. Predictive Performance of PML for Sound Quality Prediction

The PML approach demonstrated superior performance in predicting the multidimensional sound quality scores (E, P, and A-scores) compared to traditional methods. Spearman’s rank correlation tests confirmed that PML predictions achieved significantly higher correlations with the individual E-, P-, and A-scores than individual acoustic and psychoacoustic metrics (see Table 5). When evaluating the overall EPA-score prediction, the PML_EPA was found to be better than both the CRA_EPA and the GPM_EPA. While the CRA_EPA showed reasonable performance because it is a case-specific model developed using the subjective assessment results from this study, it relies on obtaining subjective data, which is often impractical for design stages. The GPM_EPA, designed as a pure-objective method without the need for subjective inputs, showed lower performance in this study (adjusted R² = 0.15). This is likely because the GPM assessment method was originally designed based on outdoor soundscapes [37], and its current formulation may not fully capture the specific noise source characteristics of indoor air-conditioning systems. However, the GPM remains valuable for building acoustics design as it offers a generalized baseline without requiring jury testing.

The significantly better performance of the PML model can be attributed to its ability to learn non-linear noise features from 227 × 227-pixel psychoacoustic heatmaps representing 0.002 s resolution evolution of loudness, sharpness, roughness, and fluctuation strength. Unlike using time-averaged stationary metrics in linear regression models (CRA and GPM), the deep learning architecture of the PML captures the complex interactions between spectral and temporal variations, offering a solution to the feature engineering problem inherent in developing pure-objective prediction models [57]. This capability makes PML the most feasible method for predicting the EPA-score, as it effectively bridges the gap between objective signal properties and subjective human perception [58], accounting the human–environmental interaction.

4.3. Implications for Occupant Well-Being Prediction

The predictive power of the PML approach extended to the assessment of occupant well-being. The standalone PML_EPA model demonstrated the highest accuracy in predicting the negative noise impacts (O1: Discomfortable, O2: Annoying, O3: Stressful, and O4: Unacceptable), surpassing both the traditional metrics and the hybrid models GPM_O1–4. Although the hybrid models (GPM_O1–4-PML) utilized the EPA-scores predicted by the PML, they yielded slightly lower goodness-of-fit compared to the standalone PHMLM_EPA model (see Table 6 and Table 7). The same domain mismatch between the GPM developed for outdoor soundscapes [37] and the PHMLM_EPA developed for indoor air-conditioned environments likely explains this result. The physics-based GPM parameters optimized for outdoor psychoacoustic relationships may conflict with the non-linear mappings learned by PHMLM_EPA from indoor soundscapes, constraining hybrid performance. Future investigation should examine GPM and PHMLM_EPA performance across both indoor and outdoor soundscapes before attempting integration. Only after domain-specific validation can data-driven predictions be effectively incorporated into generalized frameworks applicable to both environments. This indicates that simply integrating high-quality EPA predictions into the existing GPM framework, which was originally designed for different acoustic contexts, may not fully capitalize on the precision of the PML features. Therefore, the combination mechanism between data-driven predictions (PML or CRA) and generalized frameworks (GPM) requires further investigation to optimize the predictive combined effect.

In addition, the high performance of the standalone PML_EPA confirms that the method serves as a reliable assessment method for human comfort. The significance of these findings contributes directly to the field of built environment physics in studying the impact of acoustics on occupant well-being. By providing a validated, high-accuracy tool for predicting dissatisfaction and stress, the PML method empowers engineers and designers to prioritize occupant health in the acoustic design of sustainable buildings in the future [59].

4.4. Limitations and Future Work

While the PML demonstrated superior performance as a standalone model, the hybrid approach combining PML predictions with the GPM framework GPM_O1–4 yielded slightly lower goodness-of-fit for the models. This suggests that the current method of integrating data-driven inputs into the generalized model structure requires optimization. Future work should investigate more advanced fusion techniques to better leverage the collaboration between the feature extraction of PML and the generalized structure of GPM. Although some soundtracks were randomly selected from the PML training datasets, independent listening tests with new subjects generated distinct perceptual labels, validating cross-panel prediction capability. Future work should validate PML performance using completely different soundtracks from diverse HVAC systems and building types to confirm generalizability beyond the original acoustic training domain.

Moreover, the subjective perceptual responses were collected from 36 participants with a mean age of 24.1 ± 2.2 years, of whom 69.4% were male. This represents a narrow demographic range, predominantly comprising young university students. While this participant profile facilitated standardized experimental control, it limits the generalizability of well-being predictions, particularly for diverse “Office” environments where actual occupant age and gender distributions are much wider. Future research should expand participant recruitment to include a broader demographic spectrum across age (e.g., 20–65 years), gender balance, occupational diversity, and varying hearing abilities to enhance ecological validity. Furthermore, field studies in actual office settings with long-term occupant monitoring would validate model performance across real-world demographic variability. These extensions would strengthen the practical applicability of the PML method for sustainable building design across diverse occupant populations.

In addition, the PML model was trained on a specific dataset of air-conditioned sounds. Further validation is required to test its generalizability across a broader range of building types (e.g., residential, hospitals) and noise sources [60] (e.g., traffic intrusion, machinery). The AlexNet transfer learning approach may have optimized feature extraction for this particular building inventory. Future research should employ cross-dataset validation with more independent field recordings. In addition, the current study highlights the “black box” nature of ML [61]. Future research could explore Explainable-AI techniques [62] to visualize which specific regions of the psychoacoustic heatmaps contribute most to the predictions, offering deeper insights into the physical drivers of sound perception. Therefore, the impact of noise source changes on PML performance needs systematic investigation to establish stronger thresholds for practical implementation in diverse real-world scenarios. While PML demonstrated better predictive performance (adjusted R² = 0.61 for EPA-score than CRA and GPM, this is not a comparison flaw but a valid finding of method strengths. The results reflected PML’s ability to capture non-linear temporal–spectral complexity (0.002 s resolution psychoacoustic heatmaps) inherent to HVAC noise that linear methods cannot address. However, simplicity and interpretability remain critical practical advantages of CRA and GPM. CRA offers straightforward implementation using time-averaged metrics with transparent regression coefficients that enable engineering design rules. GPM provides general applicability across environments without retraining through fixed equations that facilitate rapid assessment. PML’s computational complexity represents a trade-off against CRA/GPM’s practicality for real-time building applications. Future work should validate all three methods in more engineering-practical scenarios across additional perspectives, including real-time monitoring systems, diverse HVAC configurations, long-term field studies beyond controlled jury tests, and cross-cultural perceptual validation. Hybrid GPM-PML approaches combining non-linear feature extraction with interpretable linear outputs will balance predictive power with practical utility for sustainable building acoustic design.

5. Conclusions

This study comprehensively evaluated CRA, GPM, and PML for predicting sound quality (EPA-score) and negative noise impacts on occupant well-being (O1: Discomfortable, O2: Annoying, O3: Stressful, and O4: Unacceptable) across air-conditioned built environments. PML demonstrated better performance with Spearman ρ = 0.79 (adjusted R² = 0.61) for EPA-score prediction (n = 432) than CRA (ρ = 0.58, adjusted R² = 0.32) and GPM (ρ = 0.40, adjusted R² = 0.15). Furthermore, PML also demonstrated better performance in well-being predictions (O1: adjusted R² = 0.42, O2: adjusted R² = 0.40, and O3: adjusted R² = 0.41) and showed exceptional accuracy in quiet library environments (adjusted R² = 0.53–0.62). GPM’s poor performance reflects an outdoor-to-indoor domain mismatch, while CRA’s limitations stem from static time-averaged metrics. PML’s psychoacoustic heatmaps capture critical temporal variations missed by both alternatives. This cross-method comparison identifies PML as the optimal objective assessment tool for building acoustic design, delivering reliable occupant well-being predictions without subjective testing to support health-focused sustainable building practices.