Voice is a rich modality that has garnered the interest of the research community for its potential in numerous medical applications [1,2]. Voice and speech biomarkers have recently been used for voice pathology detection [3-5], voice quality assessment [6,7], neurodegenerative disease diagnosis [8-13], mental health monitoring [14,15], cardio-respiratory condition classification [16,17], as well as automatic speech recognition (ASR) for disordered speech [18-20]. The main reason behind this interest in audio-based voice and speech biomarkers is the costliness and invasiveness of traditional voice evaluation techniques, such as laryngoscopy, stroboscopy, laryngeal electromyography, and imaging technologies such as MRI (magnetic resonance imaging) and CT (computed tomography) scans, which limit accessibility for many. Alternatively, artificial intelligence (AI)–driven audio-based medical systems pave the way for broader access to medical services for marginalized and underprivileged populations [21].

Yet, AI integration into real clinical settings is limited, in part due to the scarcity of high-quality data needed to train reliable and fair models [22]. Consequently, myriad large-scale projects are underway for the purpose of collecting extensive and representative voice audio datasets. National Institutes of Health–funded Bridge2AI is a multi-institution, large-scale project that aims to collect standardized, AI-ready, and ethically sourced voice data across various health conditions, where voice and speech samples have been collected from 442 participants [23]. Similarly, AphasiaBank (part of TalkBank Project, Carnegie Mellon University) is another National Institutes of Health–funded endeavor that has amassed multimodal data from 306 persons with aphasia [24]. Launched in 2023 and funded through 2028, SpeechDx is a global initiative dedicated to creating an extensive dataset of Alzheimer’s speech biomarkers and has recruited about 2000 participants [25]. These efforts demonstrate the general recognition of the potential of voice and speech AI in routine clinical practice.

Recently, the trustworthiness of AI systems has been a central issue for clinical integration of AI, particularly the obscurity of the decision-making processes of deep learning models to end users [22,26-29]. Known as the “black-box” problem, this issue is especially critical in medicine, where decisions directly impact patient safety. Yet, regulatory frameworks have struggled to keep pace with AI’s rapid development, leaving unresolved questions of liability and accountability [30]. Clinicians and regulators must understand how models operate, how reliable they are, and under what conditions they fail, before AI systems are integrated into clinical workflows in a meaningful way [31]. Only then can patients be guaranteed safe, high-quality medical care. On one end, there are opinions against the use of black-box models at all for high-stakes environments such as medicine [32]. While white-box, classical machine learning (ML) models such as decision trees or support vector machines trained on interpretable, hand-crafted features offer greater transparency, deep learning models consistently achieve superior performance; this is known as the interpretability-accuracy trade-off [33].

Explainable artificial intelligence (XAI) has emerged to ameliorate this challenge, aiming to make black-box models more transparent [22,29,31,34]. However, many XAI techniques have been developed for image-based domains, such as overlaying saliency maps on brain MRI or chest CT scans. In voice analysis, raw audio is often transformed into time-frequency representations such as spectrograms before being input into deep learning models. Mapping a region on a spectrogram back to an intuitively understandable auditory event (eg, a tremor on a specific syllable) is inherently more complex than identifying a visible tumor on an x-ray. The relationship between spectral features and perceived voice quality or pathology is often highly nonlinear, making clinical interpretation challenging.

Furthermore, the intuitiveness of model explanations varies substantially with respect to end users [35-37]. Computer scientists and ML researchers may be able to interpret technical visualizations such as activation maps or feature attribution plots, whereas clinicians, patients, and regulators have different knowledge bases, priorities, and constraints. A “one-size-fits-all” approach to XAI design is therefore inappropriate, especially when explainability is positioned as a pathway to increasing trust. Thereby, multidisciplinary effort in the design of XAI methods is vital for providing appropriate presentations of explanations suited for the diverse backgrounds and needs of the respective stakeholders [37,38].

In this context, this paper presents a systematic literature review of XAI methods applied to deep learning models for clinical voice and speech analysis. In this review, we use “clinical audio” and “clinical voice and speech” to refer to voice or speech data derived from individuals with a clinical condition used for a health care–relevant application or outcome, regardless of whether the recording occurred in a controlled clinical environment. The review addresses the following research questions:

The remainder of this paper is organized as follows. The rest of the Introduction section presents background for XAI concepts and different perspectives regarding the definitions of explainability and interpretability. We also outline the objectives of XAI and provide a discussion on the broad application of XAI in diverse medical domains. The Methods section describes the systematic review methodology. The Results section presents explainability approaches, explainability validation strategies, and human-centered evaluation. Then, we discuss the limitations of current approaches, stakeholder alignment, and the impact of audio representation on interpretability, upon which future directions are presented.

With recent advances in medical AI, the implementation of XAI has become increasingly critical to ensure that AI models are reliable, ethical, legally compliant, and clinically aligned, particularly in high-stakes environments such as health care. The literature contains numerous surveys discussing the role of XAI across diverse medical applications.

Several recent reviews have broadly examined the adoption of XAI in medicine, emphasizing its potential to improve transparency, foster trust, and facilitate regulatory compliance [61-65]. These works cover a variety of data modalities, including medical imaging, electronic health records (EHR), genomics, and time-series analysis. Commonly reported methods include model-agnostic techniques, most notably LIME (local interpretable model-agnostic explanations) and SHAP, alongside post hoc visualization approaches. Nonetheless, these reviews consistently highlight persistent challenges relating to faithfulness, stability, and standardized evaluation of explanations.

Domain-specific surveys further illustrate these trends. For example, van der Velden et al [66] reviewed over 200 studies applying XAI to medical imaging, noting the predominance of visualization techniques, followed by perturbation-based, textual, and example-based approaches. While visual explanations have been extensively validated, the authors stress the need for equivalent validation of textual and example-based methods. Muhammad and Bendechache [67] similarly highlight the interpretive ambiguity and perturbation sensitivity of visual methods. In the context of medical time-series data, Caterson et al [68] present a scoping review of XAI for EHR, finding feature attribution methods to be the most widely used. Salih et al [69] review XAI in cardiology for ECG (electrocardiogram) and EHR, reporting the frequent use of SHAP and Grad-CAM, followed by LIME, but also note that nearly half (47%) of the studies did not use any formal evaluation of the explanations produced. In the mental health and psychiatry domain, Joyce et al [70] reviewed XAI applied to neuroimaging, interview transcripts, and physiological data, underscoring the importance of human-centered design for producing clinically useful explanations beyond raw saliency maps. With respect to clinical audio, Chen et al [16] reviewed XAI approaches for vocal biomarker-based lung disease detection, discussing several issues, including the utility of explanations defined in terms of informativeness and user understanding as being an important criterion for evaluating explanations.

Despite the breadth of these surveys, to the best of our knowledge, no systematic reviews have comprehensively examined the use of XAI for clinical voice and speech biomarkers more broadly. This gap highlights the need for a dedicated synthesis in this domain, particularly given the unique challenges and interpretive demands of audio-based clinical decision support systems.

We conducted our systematic review according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guideline. A review protocol was prepared, but not registered before this study. Our goal was to survey papers that perform explainability or interpretability analyses of deep learning models applied to clinical voice and speech audio. A comprehensive search was carried out across IEEE Xplore, ACM, Scopus, PubMed, Web of Science, and Nature.

The following search terms were selected across three domains: explainability and interpretability, voice, speech, and acoustics, and clinical and health care context. These terms were used to create a query, which was adapted for each database. The exact search strings for each database can be found in Multimedia Appendix 1. The search was conducted across all databases on February 7, 2025, except ACM, which was searched on February 11, 2025. The search included studies published between January 1, 2015, and the date of the search. In total, 1426 records were retrieved, which were reduced to 1348 after removing 78 duplicates.

Papers were selected if they met the following criteria of being (1) focused on deep learning models applied to voice and/or speech data for health care applications, (2) applied explainability or interpretability techniques to the model, (3) reported empirical results or experimental validation, and (4) published in peer-reviewed journals or conferences.

Papers were excluded for any of the following reasons: (1) explainability and/or interpretability analysis performed for non–deep learning models; (2) used audio transcripts only (ie, purely natural language processing methods); (3) used nonclinical audio datasets; (4) not published in English; and (5) reviews, theses, or non–peer-reviewed papers.

The results of the search strategy were exported to the RAYYAN tool [76] for screening. One reviewer conducted the initial screening of titles and abstracts against the eligibility criteria, and a second reviewer independently verified the filtering decisions. Papers were excluded only if there was clear evidence in the title or abstract that the paper did not meet the criteria; otherwise, they proceeded to full-text review. The screening process resulted in 187 papers deferred for full-text review, of which 2 papers were retracted, and 2 papers were inaccessible. The final set for full-text review was 183 papers.

Full-text screening was performed by two reviewers, with deliberation and discussion used to resolve disagreements. Reasons for exclusion at this stage were documented. The final included set consisted of 30 papers that met all eligibility criteria.

Data extraction was carried out by a single reviewer using a standardized spreadsheet. A second reviewer reviewed the extracted data and deliberated on any uncertainties or discrepancies to ensure accuracy and completeness. The following information was extracted from the final set of included papers: (1) bibliographic information (authors, and year), (2) dataset information (clinical condition, number of subjects, and acoustic tasks), (3) clinical application, (4) deep learning methodology (model, hyperparameters, and training or validation strategy), (5) model performance, (6) explainability or interpretability strategy, (7) insights gained from explainability or interpretability analysis, and (8) validation or support for the explainability or interpretability results.

A formal quality and risk-of-bias assessment was conducted to evaluate the methodological robustness of prediction models estimating health-related outcomes. PROBAST+AI (Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence) [80], an extension of the original PROBAST [81] framework developed to address advances in AI- and ML-based prediction models, was used for this purpose.

A review-level PICOTS (Population; Index Model; Comparator; Outcome; Timing; Setting) framework was predefined to guide the scope and intended applicability of the assessment. The tool was applied to studies that developed or evaluated prediction models estimating health-related outcomes. Studies that did not fit these criteria were not suited for quality assessment by the PROBAST+AI tool. Each eligible study was classified as model development, model evaluation, or both. In accordance with PROBAST+AI guidance, development and evaluation components were assessed separately.

The tool evaluates four domains: (1) participants and data sources: addressing data origin, collection procedures, and representativeness; (2) predictors: examining input definition, preprocessing, and availability at the time of intended use; (3) outcome: assessing outcome definition, measurement, and timing; and (4) analysis: evaluating sample size adequacy, validation strategy, handling of missing data, risk of data leakage, and mitigation of overfitting.

Domains were rated as low, high, or unclear for development (quality concern) and evaluation (risk of bias), respectively. Applicability was assessed for the first three domains only and not for the analysis domain. The fourth domain does not include applicability considerations under the PROBAST+AI framework, as applicability refers to the assessor’s review question or intended use of a model, including the target population and setting. In accordance with PROBAST+AI guidance, if at least one domain was rated high, the overall judgment was classified as high. Detailed signaling criteria are described in Multimedia Appendix 2 [3-6,9-15,47-57,59,60,82].

The assessment was independently conducted by two reviewers, with disagreements resolved by discussion. The PICOTS framework definition and domain-specific ratings are provided in Multimedia Appendix 2.

A total of 30 studies met the inclusion and exclusion criteria. Although the search spanned publications from 2015 to 2025, all included studies were published from 2020 onward, except for a single paper. This can be explained by the relatively recent adoption of deep learning-based approaches and XAI for voice and speech in health care research. A PRISMA flow diagram summarizing the selection process is shown in Figure 1. The results are summarized in Table 2.

The included studies applied explainability methods across a broad range of voice- and speech-related health domains, including voice and structural laryngeal pathology [3-6,58,59], Parkinson disease (PD) [10,11,13], dysarthria and automatic dysarthric speech recognition [18-20,50,51], dementia [9] and Alzheimer disease (AD) detection [12], psychiatric and mental health conditions [15,48,52,55], traumatic brain injury (TBI) and focal brain lesions [47,53,54], aphasia [82], cleft palate-related hypernasality assessment [7], head and neck cancer–related intelligibility assessment [58], and autism spectrum disorder [60].

Datasets used across the included studies varied in quality, scale, language, and provenance. Most datasets were collected in controlled clinical environments, with only a single study [52] relying on remotely collected speech data. Dataset sizes ranged from approximately 15 participants to several hundred. Table 3 reports dataset sizes and tasks reported in their respective data collection protocols, not the studies included in this literature review. The datasets used represent a range of linguistic backgrounds, with English, Chinese, and Korean datasets being the most prevalent. Publicly available benchmark datasets were predominantly used in studies focusing on neurodegenerative diseases and motor speech disorders, whereas studies targeting psychiatric conditions relied exclusively on institution-specific or private datasets (eg, TORGO, UASpeech (Universal Access Speech), and PC-GITA [Parkinson Corpus – Grupo de Investigación en Telecomunicaciones Aplicadas]). Speech elicitation tasks varied across studies and included sustained vowel phonation, diadochokinesis, read speech, picture description, and free or spontaneous speech.

Explainability methods used in the included literature are summarized in Table 1. Across the 30 included studies, gradient-based saliency methods, input perturbation techniques, and model-internal representation analysis were the most frequently used explainability approaches.

Gradient-based saliency methods were reported in 7 studies [3,4,6,47-50], with Grad-CAM being the most commonly used technique, appearing in 5 studies. Grad-CAM [106] is a model-specific technique that uses gradients of convolutional feature maps to generate class-discriminative explanations as coarse (ie, non–pixel-level) heatmaps. Other gradient-based approaches, including guided backpropagation and vanilla saliency maps, were rarely used [4,50]. Vanilla saliency [107] computes pixel-level importance by calculating gradients of the model output (ie, logits) with respect to the input features, where the magnitude of the gradient at each feature indicates its importance to the model’s decision. Guided backpropagation [108] differs from vanilla saliency maps in that gradients are only propagated when both the forward activation and the backward gradient are positive, resulting in cleaner, less noisy, and more visually interpretable maps. Unlike Grad-CAM, vanilla saliency and guided backpropagation operate directly at the input level and do not provide explicit, spatially localized class-discriminative explanations.

Input perturbation-based techniques were also widely adopted across applications. SHAP was implemented in 4 studies spanning dementia detection [9], auditory verbal hallucination assessment [52], psychiatric disorder classification [51], and TBI [53]. SHAP [109] is a model-agnostic explainability method grounded in cooperative game theory that attributes to each feature a contribution score, quantifying how the presence or absence of that feature influences a model’s prediction. Ablation-based analysis was the most prevalent perturbation method, reported in 9 studies, and was applied across multiple clinical tasks [11,12,14,15,18,20,55,56,82]. Ablation studies usually involve systematic removal of specific parts of the model or input features and observing the resulting impact on model performance.

Model-internal representation analysis was reported in 12 studies, where interpretability was derived through inspection or visualization of learned internal model representations. Although primarily used as a dimensionality reduction technique, t-distributed stochastic neighbor embedding (t-SNE) [110] dominated this category, appearing in 9 studies [3,5,19,20,49,56,57,59,60], for its utility in mapping high-dimensional latent representations into low-dimensional space while preserving local neighborhood relationships. Additional representation-level analysis included inspection of learned convolutional filters [13,20], parameterized filter structures (ie, Sinc filters) [5], and dimensionality reduction of feature maps (ie, Eigen class activation maps [Eigen-CAM]) [10], providing insight into how internal model components responded to speech signals. Sinc filters are parameterized to directly model frequency bands, allowing them to more effectively emphasize formant frequencies. As Sinc filters effectively function as bandpass filters, they offer conceptually more interpretable representations compared to the abstract patterns learned by arbitrarily shaped CNN filters [5]. Unlike Grad-CAM, Eigen-CAM [111] is a gradient-free method that calculates saliency maps by taking the first principal component of flattened feature maps.

Surrogate model-based methods were less frequently used among the included literature [9,54]. Most notably, LIME [112] is a model-agnostic explainability method that learns a simple, interpretable surrogate model, such as linear models or decision trees, that locally approximates the predictions of a black-box model. An adaptation of LIME, explainable deep learning mel-frequency cepstral coefficients (xDMFCC), is another surrogate-based model designed for interpreting MFCC-based audio representations [54].

Concept-based methods, which link internal model activations to predefined, human-interpretable concepts to explain predictions in domain-relevant terms, were implemented in 2 studies [7,58].

Lastly, attention-based explanations were reported in 2 studies, where interpretability was derived either through direct visualization of attention weights [113] or through attention rollout-based aggregation of attention across layers [57].

Reported explainability methods operated over a range of input representations, which ultimately defined the form and interpretability of the resulting explanations.

Time-frequency and cepstral representations, including wideband and narrowband spectrograms, mel-spectrograms, and MFCCs, were commonly interrogated by gradient-based saliency methods [3,4,6,10,47,49,50]. In these cases, explanations were presented as heatmaps highlighting salient temporal segments and spectral bins of the input representation.

Surrogate-based methods and input perturbation techniques typically operate over cepstral representations and acoustic low-level descriptors (LLDs) [9,51]. These approaches produced explanations in the form of numerical feature importance values and ranking scores. Various audio representations were used for ablation analysis, where interpretable insight was based on tabular performance changes, following feature removal or modification [11,12,14,15,18,20,55,56,82].

Model-internal representation analyses, such as t-SNE, derived interpretability from latent representations learned by deep models, with explanations expressed as 2D projections revealing structure or separation in the latent space [3,5,19,20,49,56,57,59,60]. In a small number of studies, parameterized filter models operating directly on raw waveforms were used, where interpretability was provided through filter frequency responses identifying salient frequency bands [5,53].

Finally, concept-based approaches operated on frame-level cepstral representations, where neural activations were inspected in relation to predefined, human-interpretable semantic or clinically meaningful concepts [7,58].

The concentration of quality concerns within the analysis domain, as identified by PROBAST+AI [80], has important implications for XAI applications and the interpretability of explanations. The use of complex, high-capacity deep learning models trained and evaluated on limited clinical voice datasets, often without external validation or calibration assessment, increases the risk of overfitting to dataset-specific characteristics rather than generalizable, condition-relevant attributes.

This is especially important because post hoc explainability methods are contingent on the model’s internal representations and input-output relationships. Explanations derived from overfitted or insufficiently validated models may appear coherent while reflecting confounding artifacts or spurious correlations [32,72]. Thus, the credibility of XAI methods does not rest solely on the faithfulness of the explainability technique, but also on the methodological rigor of the development and evaluation of the underlying predictive models.

These considerations are particularly relevant for the analysis of the domain-specific explanation patterns. Although recurring themes were observed across the clinical domains, the risk of bias of the underlying models and circular validation across studies due to the use of the same dataset should motivate the reader to view the findings or explanations with caution, as they may reflect data-specific characteristics rather than generalizable disease-related insights. Accordingly, the purpose of presenting domain-specific explanation patterns is not to establish definitive clinical explanatory signatures, but rather to characterize the current landscape of explainability practice in voice and speech AI and to motivate more rigorous, validated, and clinically grounded approaches in future work.

Although current XAI methods provide preliminary insight into the inner workings of clinical audio models, the interpretation of explainability outputs is rarely subjected to rigorous validation. Many studies rely on a limited number of illustrative examples to interpret local explanations (eg, saliency maps or attention weights) without quantitative assessment or statistical analysis [3,4,6,48,50,57]. This increases the risk of overinterpretation such that visually compelling or anecdotal explanations are inferred to be clinically meaningful despite limited evidence of generalizability.

Some studies attempt to contextualize explanation outputs by comparison with established clinical or acoustic knowledge [9,10,12,48,50-52,54]. While such comparisons may enhance face validity, they do not guarantee faithfulness of the explanation to the model’s true decision process. Prior work has demonstrated, for example, that attention weights are not inherently faithful indicators of feature importance and may exhibit weak or inconsistent correspondence with gradient-based relevance measures, leading to potentially misleading interpretations [114-116]. Interpreting explanations primarily through the lens of existing medical literature may therefore introduce confirmation bias, whereby explanations that align with prior expectations are accepted uncritically while alternative patterns are overlooked [35].

These concerns are reinforced by the fact that most reviewed studies [3,6,9,10,12,15,18,47,48,50-54] did not perform any form of quantitative explainability validation. In the absence of systematic evaluation, explanatory interpretations are more vulnerable to confirmation bias and may overstate their clinical relevance. This pattern might be indicative of underlying issues relevant to clinical voice and speech research. The scarcity of qualitative explanation validation is possibly driven by the high cost and, subsequently, the scarcity of annotated, high-quality clinical voice and speech datasets. The same reason might also explain the high risk of bias for most included studies, even for otherwise methodologically sound ones.

The reviewed studies suggest an inherent trade-off between model complexity, input representation, and explainability in clinical audio-based AI systems [32]. In particular, the interpretability of model explanations is closely linked to the degree of semantic transparency in the input features. Audio representation in the literature included raw audio [49], low-level acoustic descriptors (eg, jitter, shimmer, and harmonic-to-noise ratio) [51,60], raw time-frequency representations (eg, spectrograms and mel-spectrograms) [6,48,50,82], coefficient-based features (eg, MFCCs) [7,11,56,82], and deep neural network–transformer-based embeddings (eg, i-vectors, wav2vec, and HuBERT [Hidden-Unit Bidirectional Encoder Representations From Transformers]) [14,18,19,60].

Models trained on manually extracted LLDs offer the highest degree of interpretability, as these features are directly related to established clinical biomarkers or constructs. Explanations from models trained on LLDs can be relatively easily understood using feature attribution methods such as SHAP or LIME. This alignment between model explanations and clinical knowledge contributes to the popularity of such methods in the medical AI domain.

While MFCCs are also hand-engineered features, they present interpretability challenges due to their abstract and decorrelated nature [117]. The coefficients do not correspond directly to intuitive phonetic or physiological phenomena. Although recent techniques such as xDMFCC attempt to provide coefficient-level explanations across time, and Tracey et al [117] sought to demystify MFCCs as vocal biomarkers by correlating them with known LLDs, the clinical relevance of these explanations remains limited. Despite efforts to position MFCCs as vocal biomarkers, they largely obscure the internal workings of deep models and contribute to the opacity of clinical audio systems [118].

While raw audio exhibits energy and temporal information, spectrograms offer a more interpretable alternative. As visual time-frequency representations, they capture dynamic changes in energy that are often physically and clinically meaningful, such as vocal formants or pauses. Trained speech pathologists and clinicians can interpret spectrograms directly, making visualization-based explanations (eg, saliency maps) more accessible and potentially clinically actionable compared to MFCCs. Although visual explanations of spectrograms might inform about the time segment most important for the detection of a condition and can highlight relevant frequency bands, more granular visual explanations are typically not feasible.

Finally, state-of-the-art systems increasingly rely on transformer-based embeddings such as wav2vec and HuBERT, which yield task-optimized, high-dimensional representations learned from raw waveforms. While these embeddings deliver substantial performance gains, they are the most difficult to interpret due to high-dimensional, nonlinear abstraction from both low-level acoustic features and clinically grounded descriptors. As a result, explanations tend to be less granular, less transparent, and harder to align with clinical reasoning.

A major theme in the literature is the misalignment of the explanation form or modality and the needs of the end users or stakeholders. All included studies produce explanations targeted for highly technical audiences (ie, AI researchers and developers), without taking into consideration the interpretive frameworks of clinicians, patients, regulators, and policymakers. While such explanations offer insight to AI developers and researchers, critical for understanding model behavior, enhancing performance, and ensuring reliability, these explanations are often too technical, abstract, or detached from domain-specific language to be directly actionable in clinical decision-making. Figure 2 highlights the central challenge of stakeholder–explanation misalignment in clinical voice and speech audio AI, illustrating how diverse stakeholders hold distinct expectations and informational needs, and emphasizing the necessity of tailoring explanation strategies accordingly.

Explanations that focus on algorithmic mechanisms poorly convey information in terms of established diagnostic criteria or clinical reasoning processes [35]. Patients, with their highly limited technical and clinical knowledge, are less likely to draw relevant insights from such technical explanation modalities such as activation maps and feature attribution plots [119]. Similarly, regulators and policymakers prefer explanations that offer transparent, auditable decision pathways to assess compliance, fairness, and accountability [120].

This gap is partly a result of explanation design being driven primarily by XAI method availability rather than user requirements [121]. Most of the reviewed literature adapts generic explainability techniques to clinical audio tasks without taking into consideration the information needs, domain expertise, or cognitive constraints of their target users. This highlights the imperative need for human-centered, context-aware design of clinical audio explainability methods in which concerned stakeholders participate and provide valuable input for suitable explanation modalities and content.

In this work, we discussed several limitations of current methods and identified opportunities for advancing interpretability and explainability in clinical audio-based deep learning systems. Accordingly, this section outlines future directions and recommendations for advancing XAI specifically for clinical audio applications, as summarized in Figure 3.

Although the literature used various explainability strategies, several techniques that are well established in related domains have rarely been adapted for clinical audio-based systems. For instance, example-based and counterexample-based explainability methods are widely used in audio-based emotion recognition [73] to generate sonified explanations. Similarly, in the text-to-audio domain, AudioGenX quantifies the importance of textual tokens corresponding to generated audio using factual and counterfactual techniques [16]. Integrating these approaches into clinical voice and speech analysis can enhance explanation fidelity and stakeholder interpretability.

Moreover, many of the XAI techniques surveyed in this review are originally developed for image or tabular data, underscoring the need for domain-specific approaches tailored to the temporal-spectral nature of audio signals and the heterogeneous manifestations of speech and voice disorders. This is especially important given the perceptual nature of audio. These methods should aim to map abstract representations (eg, deep transformer-based embeddings) to established clinical constructs, thereby bridging the gap between model outputs and actionable clinical insight. For example, CoughLIME (Cough Local Interpretable Model-Agnostic Explanations) [122] extended LIME to generate sonified explanations for COVID-19 cough analysis. Future work should aim to develop sonified XAI that aligns with the perceptive ability of clinicians (eg, speech-language pathologists). Recent progress in large audio language models is garnering interest due to chain-of-thought reasoning and their capability to identify environmental sounds, speech characteristics, and respiratory and heart sounds [123,124]. These models can be used to extract salient acoustic features and present them in structured, report-style textual summaries through a stakeholder-friendly interface.

Our review also identified a lack of rigorous validation of XAI explanations. Future work should use both quantitative measures (eg, fidelity, sensitivity, perturbation-based testing, and cross-dataset explanation consistency) and qualitative, human-centered evaluation strategies (eg, expert annotation comparison and interrater agreement). For example, a recent work introduces a frequency band perturbation framework for quantitatively evaluating the faithfulness of various XAI techniques [125].

Additionally, the misalignment of explanation formats with stakeholder needs highlights the importance of iterative, collaborative design processes in which AI developers engage clinicians, patients, and regulators throughout system development. Such collaboration can help ensure that explanations are understandable, clinically relevant, and operationally feasible, while supporting auditability, transparency, and accountability. For instance, the work by Pizzimenti et al [126] describes a Delphi study aimed at unifying and standardizing vocal biomarker research where clinicians, statisticians, audio signal processing experts, AI researchers, and ethicists are involved in this endeavor. Such efforts highlight the importance of structured, interdisciplinary validation frameworks for clinical audio research.

Finally, throughout this systematic review, we encountered substantial difficulty in quantifying the degree of explainability and practical utility of reported methods, particularly across different stakeholder groups. Consistent with prior findings [44], only a small number of studies formally evaluated the effectiveness of XAI in clinical settings. Future work should therefore develop composite evaluation frameworks that integrate objective indicators (eg, performance improvement and error detection) with subjective measures (eg, perceived clarity, trust, and usability), enabling comparison of explainability methods in terms of both model alignment and real-world clinical impact. This is achieved by addressing limitations identified under PROBAST+AI, such that future work prioritizes study designs with representative cohorts, sufficient sample sizes, and appropriate model evaluation of diagnostic and prognostic modeling. Such practices will ensure the development of methodologically sound models, such that clinical outcomes derived from explainability methods are of high quality and clinical value, thus increasing the trustworthiness of AI in clinical practice.

This review has several limitations that should be considered when interpreting the findings. Study selection and data extraction were conducted by a single reviewer. Although established PRISMA procedures and predefined inclusion criteria were followed, the absence of a second independent screening may increase the risk of selection bias or missed studies.

Second, conclusions regarding domain-specific explanation patterns are constrained by the methodological quality of the underlying prediction models. A substantial proportion of studies exhibited a high risk of bias, particularly in model development and evaluation. In addition, several studies relied on repeated use of the same benchmark datasets across dysarthria, PD, and AD domains, raising the risk of circular validation and inflating apparent consistency of explanation patterns. The domain-specific synthesis presented in this review is therefore intended to characterize current practice rather than establish definitive clinical explanatory signatures.

Finally, given the rapid growth of this field, studies published after the search period (February 2025) may provide additional insights and are not reflected in this review.

In this systematic review, we presented current practices of explainability and interpretability for deep learning–based voice and speech analysis in clinical care. Across 30 eligible studies, we identified a diverse set of explainability methods, which we organized into commonly adopted categories. Our findings indicate that, although explainability techniques are increasingly applied across a wide range of clinical speech and voice applications, their use is largely exploratory and rarely supported by rigorous validation. Explanations were predominantly assessed through qualitative interpretation, with limited evaluation of faithfulness, robustness, or consistency across datasets, and no explicit human-in-the-loop assessment involving clinical or regulatory stakeholders. Additionally, the quality of underlying models limits the validity of the reported explanation patterns and, subsequently, the applicability of these models for real-world clinical applications. These findings highlight the need for domain-specific, clinically grounded explainability methods, standardized validation protocols, and stakeholder-aware explanation design to support the safe and effective integration of voice and speech AI into clinical practice.