The production of human voice and speech is an intricate process that involves the coordination of respiration, phonation, articulation, resonation and prosody (1, 2) and depends on the collaboration of several motor and cognitive processes (2, 3) and more than 100 muscles (4).¹ These processes are directly impacted by human physiological features. For example, the length and shape or human vocal tract and the dynamic configuration of articulatory organs can directly affect sound and speech production (7). Voice and speech enable humans to convey their moods and assert their personality (8), communicate with their environment, express emotions, and detail thoughts (9–11). Speech and voice, through both their physiological basis and social use, shape auditory identity, serving as an “invisible business card” presented to the world (12).

In recent years, these characteristics have led to the popularization of using vocal biometrics for tasks like voice authentication, speaker detection, and forensic speaker recognition (13). The uniqueness of voice and speech has even been compared to that of fingerprints (14), although some authors have contested this comparison (15). Recording voice and speech is quick, inexpensive, non-invasive, and can be done remotely (14–17). Furthermore, recognizing people using their voice and being recorded are familiar processes, especially as microphones have become more ubiquitous in modern day life (18).

Aided by the developments in the field of voice biometrics and the successes of data-driven approaches for consumer speech applications, there has been a recent surge in interest in leveraging human voice and speech as sources of data for healthcare. Voice and speech biomarkers, derived from the nuances in voice and speech, represent a powerful yet underused resource in healthcare (19). Traditionally, biomarkers have been described as “a defined characteristic that is measured as an indicator of normal biological processes, pathogenic processes, or biological responses to an exposure or intervention, including therapeutic interventions” (20). Biomarkers can have different putative applications, ranging from prediction, diagnostic, or monitoring of a condition or a treatment, to the study of pharmacodynamic responses to medications. Voice biomarkers are a part of the new frontier known as digital biomarkers (21), where behavioural and physiological data is collected using digital devices to aid in diagnosis, monitoring, and treatment of various health conditions (22, 23). Although the term digital biomarker has recently been under some scrutiny for being misleading, unclear, and ill-defined (24–26), no other common term has emerged or gained consensus among scholars. Voice biomarkers are defined as distinct vocal attributes, alone or in combination, that have been scientifically validated as indicators of clinical outcomes (27, 28). They can be used for symptom detection and monitoring, diagnosis of conditions, screening potential health issues, perceived health status monitoring, and to monitor disease progression (29). Similar to its use as biometric information, voice biomarkers have the advantage of being easy and relatively inexpensive to collect and preserve in databases (30, 31).

Even though the idea of using voice as a biomarker of health has been explored for decades (32), recent developments in artificial intelligence (AI), machine learning (ML), and natural language processing (NLP) have enabled fast, precise analysis of voice data, making it possible to envision the scalable development of voice biomarkers analysis (33–35). Voice biomarkers have already demonstrated their potential in the diagnosis of diseases and conditions such as Alzheimer's disease (36), Parkinson's disease (37), diabetes, developmental problems such as autism (38), and various mental health problems such as depression, stress, and suicidal ideation (39, 40). However, there are a limited number of voice datasets meant for clinical use (41). These datasets are often incomplete, and due to this fact researchers often lack the capacity to consider implications of diversity and privacy control (42) or properly train algorithms with them (27). This has led to calls to curate large, scalable, standardized databases to use for research on voice and speech biomarkers (34, 43, 44).

While work has been done on the ethical, legal, and social implications (ELSI) of using voice biometrics (45–48), the collection of voice and speech for clinical use requires a deeper understanding of the ELSI specific to this nascent use (44, 49). This scoping review aims to present the state of the literature on the ELSI that arise from the collection and use of voice and speech data in healthcare, while also adding to the growing literature on the ELSI of digital biomarkers (50–53). Moreover, it aims to identify how the collection of voice and speech data for medical and clinical use can redefine some of the ELSI related to digital biomarkers, how it fits with current understanding of the ELSI related to biomedical and clinical data, and to address the gaps in the existing literature and potential areas of concerns. The aim of this analysis is to contribute to the responsible development and use of voice and speech biomarkers.

A scoping review was realized to appraise and determine the volume of literature and studies available on the ELSI of the use of voice as a biomarker of health (54, 55). A scoping review approach was selected because it allows for a comprehensive mapping of the diverse studies and disciplines addressing the evolving challenges posed by voice data in healthcare. Rigor was ensured by engaging in a predefined and comprehensive search strategy (56, 57). This scoping review was conducted in 3 phases. A first phase (Phase 1) consisted in finding articles through a search on databases, a second phase (Phase 2) was added where articles were found through snowball research from Phase 1, and a third phase (Phase 3) was conducted to review new literature. Each phase, and its associated steps, are described below. The charting of the data and the collating, summarizing and reporting of results are presented in the Results section.

Results are presented following the different themes of the coding. Each of the three types of implications (ethical, social, and legal) are examined in two ways. The first category, “conventional implications”, covers issues arising from the use of voice and speech data in healthcare that parallel those linked to other biometric data, such as gait or ocular movement analysis (63, 64), and defines where voice and speech are treated as continuous with existing health data ethics discussions. The second category, “modality-specific implications”, addresses ELSI that require refinement or revision due to the unique characteristics of voice and speech data, and defines where voice and speech data are challenging conversional frames.

As noted, not all of the included articles were specifically related to voice biomarker research or the use of voice biomarkers in healthcare. Although they addressed voice biomarker research and use, some articles dealt with digital biomarker research more generally, the use of smartphones or sensors in healthcare, or new trends in the treatment of a specific disease. Contrasting these implications with those brought on specifically by voice and speech data makes it possible to recognize the distinct risks and responsibilities associated with voice, while maintaining coherence across wider conversations on health data governance. In each of those categories, the implications are presented in order of how frequently they were mentioned by authors.

This review examined the ethical, legal, and social implications (ELSI) of voice and speech data in healthcare, focusing on whether their use presents novel or exceptional challenges. Many of the identified ELSI are not unique to voice and speech data. These concerns have been extensively discussed in the context of biomedical technologies and other types of biomarkers (124–127). However, the distinctive features of voice and speech require reinterpreting existing frameworks.

Tables 2–4 present a side-by-side to illustrate how the use of voice and speech data in healthcare both aligns with conventional implications or brings out modality-specific implications. Ethical concerns such as privacy, consent, and data quality persist, but take on new dimensions due to the technical and inferential characteristics of voice. Social implications reveal how structural biases are amplified through underrepresentation and linguistic exclusion, while also introducing risks of exploitation specific to audio data. Legally, although some regulatory gaps are familiar, voice data introduces new complications in classification, protection, and admissibility. The divergence in these columns shows where existing frameworks are strained and where entirely new considerations arise. This type of analysis is important because the unique characteristics of voice introduce challenges not captured by existing discussions of other health data. Integrating these concerns into broader ELSI conversations ensures that the distinct risks and responsibilities tied to voice are recognized, while maintaining coherence across debates about health data governance.

A key observation is the interdependence of ethical, legal, and social implications in this domain. For example, ensuring diversity in voice datasets is more complex than in other biomarker domains (27). Although diversity is essential for reducing bias, building representative voice datasets is resource-intensive and logistically difficult. This challenge is compounded by fragmented legal protections and inconsistent privacy regulations. Several articles question whether existing national and international regulations adequately govern voice biomarker development. While GDPR and HIPAA provide some safeguards to patients and system users (72, 88, 97, 116, 121), their limitations are evident, especially since many companies rely on them as default governance models (109). National and international regulatory frameworks must be adapted to address the specific risks posed by new biomarkers, including voice-based digital biomarkers (71, 78). Regulatory responses must keep pace with rapid industry development and clarify legal responsibility (71, 91). Beneath these regulatory challenges lie unresolved technical uncertainties. While outside the scope of this review, these gaps limit a comprehensive understanding. Nevertheless, the recurring mention of technical issues (such as inconsistent labeling and processing) shows their influence on both ELSI and mitigation strategies.

Addressing the intersection of ethical, legal, social, and technical implications requires transdisciplinary and cross-sector engagement. As shown, the reviewed literature spans voice technology, machine learning, AI, and digital health instrumentation. Development of AI tools must incorporate insights from adjacent disciplines and from stakeholders (e.g., patients, clinicians, and advocates) throughout the design pipeline and medical AI lifecycle (50, 71, 91). The field of bioethics can offer guidance in curating these discussions, ensuring stakeholder representation and participation, and generating context-specific guidelines that go beyond overarching principles and general claims about privacy, informed consent, and data security (50).

Few articles included in this review addressed the private sector, likely due to the methodological constraints of the search strategy and inclusion criteria. Those that did, however, tended to highlight the widening gap between academic researchers and private industry (82, 91). This was mostly due, as reported in the results, to the scarcity of large, diverse, high-quality datasets (36, 37, 40, 128, 129). This limitation motivated initiatives like the Bridge2AI-Voice consortium (130), which aims to develop an ethically sourced database of thousands of voice samples linked to health metadata (34, 49, 131). Historically, large-scale data collection projects from healthcare institutions have contributed to advance considerations for the responsible stewardship and governance of data (132, 133), as well as providing viable solutions for the lack or inaccessibility of relevant data for research teams (134). But, as shown by (91), some researchers seek access to non-clinical voice recordings, such as those held by tech companies or in public repositories. These datasets, while abundant, often lack proper curation and metadata annotation needed for healthcare applications (135, 136), something that appears to be lacking in private datasets (132). Collaboration between academic and industry sectors remains valuable, especially as private actors lead technological development in this domain (137). Yet, persistent communication gaps and contradictory incentive schemes across sectors impede alignment and slow field advancement (67).

The literature remains sparse on ELSI specific to voice and speech data in healthcare. Though the field is relatively new, having emerged in the early 2000s (138), recent years show a growing engagement with ethical, legal, and social dimensions. Despite this trend, empirical studies and in-depth conceptual analyses are still lacking. Most reviews mention ELSI superficially without probing underlying assumptions or stakeholder perspectives. This gap underscores the need for further research, particularly qualitative work exploring contributors' views on privacy, consent, data usage, and trust. Understanding the motivations and concerns of data subjects is critical for ethical implementation.

Most publications on voice and speech data in healthcare originate from researchers based in the Global North (see Figure 3). This reflects the geographic concentration of datasets, infrastructure, and funding, as well as restrictive international privacy regulations that complicate cross-border data sharing (71, 88, 97). These conditions limit international collaboration and prevent research teams in low- and middle-income countries (LMICs) from accessing or contributing to existing datasets. The same imbalance holds for industry: most companies developing voice AI tools are based in the West (50). While this geographic clustering may ease local industry–academia collaboration, it undermines global inclusivity. LMICs often lack the technical, financial, and institutional capacity to generate or refine voice data (132, 139). This exclusion means that a large portion of the global population is left out of both the development and the benefits of voice biomarker technologies. Moreover, research priorities and disease targets are shaped by Global North interests. The COVID-19 pandemic exemplifies this dynamic: while voice AI tools for cough analysis showed promise, their global utility was undermined by training data biases skewed toward wealthier regions (140, 141). This case underscores how big data production mirrors (and reinforces) existing global inequities (142, 143).

As mentioned, one of the central questions in this review is whether the use of voice and speech data in healthcare constitutes a form of exceptionalism. Rather than endorsing voice biomarker exceptionalism (i.e., a claim that this domain requires entirely new ethical frameworks), this paper argues for contextualism. Contextualism, as drawn from bioethical debates in genomics and neuroethics, rejects binary categories (exceptional vs. non-exceptional) and instead emphasizes that ethical implications vary according to the specific features of the data, its use context, and the sociotechnical systems surrounding it (144–147). Voice data, in this framing, is neither wholly new nor entirely analogous to prior biomarker domains. It presents a “fundamental duality” (145), carrying specific risks (e.g., such as the inferential excesses of emotional AI, the misclassification of linguistic minorities, or the covert extraction of demographic attributes), but these risks are best addressed by adapting existing frameworks to new contexts, rather than discarding them. The ELSI mentioned in this review are not necessarily unique to voice and speech biomarkers, but they do take on different meanings and understandings and require new mitigation solutions given the unique characteristics of these biomarkers. The question is, however, interesting and even important to ask.

These contributions support a contextualist turn: ethical frameworks should not presume novelty but must remain responsive to the specific risks and affordances of voice AI. In this light, the development of voice and speech biomarkers should be understood as part of a broader evolution in digital health, comparable to past transitions seen in genomics or neuroimaging (148–151). Lessons from those domains (about oversight gaps, translational bottlenecks, and stakeholder exclusion) are not just relevant, but necessary guides for the governance of voice data in health.

This review has several limitations that reflect both structural constraints in the field and methodological boundaries inherent to scoping reviews. First, although this review refers to “voice and speech data” as a combined category, the technical distinction between voice and speech remains imprecise in the reviewed literature. Voice generally refers to the acoustic signal produced by the vocal tract, while speech encompasses the linguistic content of that signal. These modalities are often treated interchangeably, but raise distinct ethical, legal, and social concerns depending on which is being collected or analyzed. This lack of definitional consensus limits the granularity with which ELSI concerns can be differentiated and analyzed (43). Second, private industry perspectives are significantly underrepresented in the reviewed material. While it is common for industry work to remain unpublished or appear in grey literature, the academic databases consulted (e.g., MEDLINE, Web of Science, EMBASE, IEEE) primarily index peer-reviewed sources. As a result, commercial practices and proprietary standards (especially around data acquisition, model training, and consent) are largely absent from this analysis. This is only compelled by the current state of voice AI health-tech sector, which currently offers limited disclosure and transparency on their public-facing websites about their policies (152) and of their technology (153). Third, while many included articles mention ELSI, their engagement is often brief or speculative rather than empirical or conceptually grounded. This limits the capacity of the review to draw out structured, well-theorized distinctions across domains, particularly in areas such as incidental findings, third-party data capture, or intergenerational data ethics. Fifth, as a scoping review, this article maps the breadth of existing literature rather than evaluating its quality or methodological robustness. This is appropriate for identifying gaps and setting research agendas but limits the evidentiary strength for normative claims or policy recommendations.

To our knowledge, this is the first scoping review examining the ELSI associated with the collection and use of voice and speech data in healthcare. Reviewing literature is an important initial step in conducting ethical foresight. It allows to identify issues and implications that are under consideration, uncover those which are not yet acknowledged, and highlight those for which further attention is required. The emergence of voice and speech-based biomarkers reframes how voice is understood; not only as expression, but as a data source conveying cognitive, physiological, and affective information. Voice and speech have long served as instruments of identity, expression, and social interaction. They are now increasingly central not only to human communication, but also to human-machine interaction. Voice also operates as a metaphor for representation, agency, and political visibility (154). With the advent of voice biomarkers, voice is no longer just symbolic, it is also a computational object for encoded cognitive, neuromuscular, and physiological signals (155, 156).

This review also identifies several areas where additional research and policy development are urgently needed. In particular, the collection and use of voice and speech data in health contexts raise distinctive ethical concerns due to their capacity to reveal sensitive, inferential health information. The future trajectory of voice biomarker research remains uncertain, especially given the difficulty of building datasets that are trustworthy, clinically meaningful, and shareable across contexts. This review aims to support the ethical development, deployment, and governance of voice and speech data practices in healthcare and related domains.