Inter-rater reliability—which measures the consistency of evaluations between different raters—has generally been reported as high for the CAPE-V across various languages and vocal attributes. For instance, the original English version demonstrated strong inter-rater reliability for overall severity (Karnell et al., 2007). Adaptations of the CAPE-V have also shown high inter-rater reliability in certain cases. The European Portuguese version (Jesus et al., 2009) exhibited excellent inter-rater reliability for breathiness and loudness, while the Kannada version (Gunjawate et al., 2020) displayed consistently high reliability across all attributes. Similarly, the Italian adaptation (Mozzanica et al., 2014) reported high inter-rater reliability for overall severity and breathiness, and the Turkish version (Özcebe et al., 2019) demonstrated strong reliability for overall severity and loudness. The Japanese adaptation (Kondo et al., 2021) also showed high inter-rater reliability across all attributes. Finally, the European French version (Pommée, Shanks, Morsomme, Michel, & Verduyckt, 2024) demonstrated good reliability for overall severity.

However, not all attributes performed consistently across studies. The original English version exhibited low inter-rater reliability for strain, pitch, and loudness (Zraick et al., 2011). Likewise, the Persian (Salary Majd et al., 2014), Brazilian Portuguese (Behlau et al., 2022), and Turkish versions (Ertan-Schlüter et al., 2020) demonstrated low inter-rater reliability for pitch and loudness. Roughness also showed variability, with lower inter-rater reliability reported in the English (Kelchner et al., 2010; Zraick et al., 2011), Brazilian Portuguese (Behlau et al., 2022), Tamil (Venkatraman et al., 2022), and Hindi (Joshi et al., 2020) versions. The Tamil version further showed lower inter-rater reliability for pitch and strain. Additionally, the bilingual Catalan/Spanish version (Calaf & Garcia-Quintana, 2024), inter-rater reliability for breathiness and loudness was notably low, further highlighting the challenge of achieving consistency in perceptual voice assessments. Similarly, the European French version (Pommée, Shanks, et al., 2024) showed low inter-rater reliability for most parameters except overall severity,

Intra-rater reliability—which measures the consistency of the same evaluator over time—has generally been reported as high for the CAPE-V across different languages and vocal attributes, though some variability persists. For instance, the original English version demonstrated strong intra-rater reliability for overall severity (Karnell et al., 2007). Subsequent adaptations, such as the Italian version (Mozzanica et al., 2014), consistently showed high intra-rater reliability across attributes like overall severity, roughness, and breathiness. Similarly, the Spanish (Núñez-Batalla et al., 2015), Turkish (Özcebe et al., 2019), Japanese (Kondo et al., 2021), and Kannada versions (Gunjawate et al., 2020) displayed strong intra-rater reliability across most attributes. Finally, the European French version (Pommée, Shanks, et al., 2024) also demonstrated good intra-rater reliability for overall severity, strain, and pitch.

However, some studies have reported lower intra-rater reliability in specific dimensions. The English version exhibited lower intra-rater reliability for strain (Zraick et al., 2011). Likewise, the Persian version (Salary Majd et al., 2014) showed reduced intra-rater reliability for pitch and loudness, while the Brazilian Portuguese version (Behlau et al., 2022) also reported lower intra-rater reliability for breathiness and loudness. The Tamil version (Venkatraman et al., 2022) demonstrated moderate intra-rater reliability for pitch and strain, while the bilingual Catalan/Spanish version (Calaf & Garcia-Quintana, 2024) reported particularly low intra-rater reliability for loudness, highlighting the ongoing need for tools and strategies that can help evaluators achieve more consistent auditory-perceptual judgments.

The variability in auditory-perceptual judgments underscores the need for further standardization and the development of reference materials to improve reliability. Efforts such as the simulations created by the University of Wisconsin-Madison (Connor, Bless, Dardis, & Vinney, 2008), the Perceptual Voice Qualities Database (Walden, 2022), and the recent anchor and training sample set for auditory-perceptual voice evaluation (Labaere, De Bodt, & Van Nuffelen, 2023) aim to reduce this variability by providing structured materials that help standardize perceptual judgments.

The University of Wisconsin-Madison simulations consist of 45 cases, each including CAPE-V sentence samples, conversational speech, and sustained vowels /a/ and /i/. Each case also provides a detailed patient history, which can be used separately for class discussions or reflections on how medical conditions or symptoms might influence a patient's voice and what further assessments are needed for a complete voice evaluation. While the quality of this material is excellent, the application of the CAPE-V remains on paper, is not digitized, does not collect data, and offers limited feedback, as it only allows users to compare their evaluations to a single reference.

The Perceptual Voice Qualities Database (Walden, 2022) offers a greater potential, with each sample being evaluated twice by 3–4 evaluators. However, the raw data from these evaluations require processing to be effectively used in training programs. Unlike the simulations, this resource presents a larger dataset and the possibility of deeper analysis, but its full integration into educational settings remains dependent on further development to transform the raw data into actionable training materials.

The anchor and training sample set developed by Labaere et al. (2023) uses the GRBAS scale (Hirano, 1981), which, while still widely used, is now superseded by the standardized CAPE-V. The CAPE-V is internationally recognized for its standardized protocol for sample collection, refined vocal attributes, and use of a visual analog scale (VAS), allowing parametric statistical tests and improving reliability across clinical settings. A strength of Labaere et al.’s material is the strict reliability criterion for anchor samples, requiring 90% expert agreement on at least one attribute. However, relying on the GRBAS scale limits its alignment with more up-to-date practices like the CAPE-V, which offers better standardization in auditory-perceptual evaluations and allows for a more nuanced assessment than a four-category scale.

In response to these challenges, the All-Voiced app (Calaf, 2024b) was developed to provide a more accessible, scalable, and data-driven platform for both training and research in voice evaluation. The app, which is currently free to use, focuses on improving the consistency and accuracy of auditory-perceptual ratings by allowing users to practice their evaluations, receive real-time feedback, and compare their assessments with those of previous evaluators. This feedback system, combined with the app's interactive features, is designed to enhance reliability in both inter-rater and intra-rater assessments, addressing key gaps in the field.

The All-Voiced app is designed not only as a training tool but also as a research platform, with the goal of gathering large volumes of perceptual data from users with varying levels of expertise. This data collection is intended to support ongoing research aimed at refining auditory-perceptual voice evaluation techniques and identifying patterns that could inform the development of more targeted training materials and clinical tools in the future.

The primary aim of this study is to present preliminary findings on the user profiles and perceptual judgments collected during the first 75 days following the launch of the latest app All-Voiced app version. Specifically, this study analyzes the user registration patterns and voice evaluations of the most frequently assessed samples, and highlights the potential of the app to provide valuable insights into the perception of vocal qualities as more data is collected over time.

In this first phase of the app implementation, 120 vocal samples were selected from the Perceptual Voice Qualities Database (PVQD) (Walden, 2022) for use within the app. This selection was carefully curated to ensure a balanced representation of age groups, gender, diagnoses, and varying severities of dysphonia, providing a comprehensive dataset for auditory-perceptual evaluations. To standardize the listening conditions, all vocal samples were normalized to 70dB, ensuring a consistent loudness across evaluations. Additionally, the sociodemographic labels were standardized—for example, unifying variations such as “f”, “F”, “female”, and “Female” under a single label for gender, and harmonizing diagnosis terms to improve consistency and comparability across different groups.

To ensure accessibility across different regions, the app supports multiple languages. All versions of the CAPE-V available within the app have been included with permission from the authors and copyright holders. The currently available CAPE-V versions are English (Kempster et al., 2009), Spanish (Calaf & Garcia-Quintana, 2024), Catalan (Calaf & Garcia-Quintana, 2024), French (Pommée, Mbagira, et al., 2024), and Japanese (Kondo et al., 2021). The entire app has been translated into these languages, ensuring comprehensive accessibility. To ensure cultural and linguistic appropriateness, the app includes a feedback form in the footer, inviting users to contribute to improving the language adaptations.

At the time of writing this manuscript, negotiations are ongoing with the authors to include additional language adaptations of the CAPE-V.

The statistical analyses and visualizations were conducted using Python (version 3.10.12) in a Google Colab environment. Key libraries included pandas (version 2.2.2) for data manipulation, matplotlib (version 3.8.0) and seaborn (version 0.13.2) for visualizations.

For the participants’ sociodemographic data, descriptive statistics were calculated to summarize the distribution of age, gender pronouns, country of residence, language preference, profession, and voice competence levels. The results are presented in counts and percentages.

To analyze the gathered auditory-perceptual evaluations, scatter plots and box plots (Figs. 1–4) were created to compare the evaluation results from All-Voiced users across different voice competence levels (beginner, intermediate, advanced, and expert) with those of the Perceptual Voice Quality Database (PVQD) by Walden (2022) for standard perceptual attributes such as Overall Severity, Roughness, Breathiness, and Strain.

Figure 1.

Distribution of overall severity scores for samples with 10 or more unique all-voiced raters, showing expertise levels of all-voiced participants and PVQD data. Note. PVQD=Perceptual Voice Qualities Database (Walden, 2022).

(0.11MB).

Figure 2.

Distribution of roughness scores for samples with 10 or more unique all-voiced raters, showing expertise levels of all-voiced participants and PVQD data. Note. PVQD=Perceptual Voice Qualities Database (Walden, 2022).

(0.11MB).

Figure 3.

Distribution of breathiness scores for samples with 10 or more unique all-voiced raters, showing expertise levels of all-voiced participants and PVQD data. Note. PVQD=Perceptual Voice Qualities Database (Walden, 2022).

(0.11MB).

Figure 4.

Distribution of strain scores for samples with 10 or more unique all-voiced raters, showing expertise levels of all-voiced participants and PVQD data. Note. PVQD=Perceptual Voice Qualities Database (Walden, 2022).

(0.11MB).

In this comparison, all available evaluations from the PVQD database were utilized, as each rater provided two independent assessments per sample. This approach ensured the maximum use of the limited data available from this source (3–4 raters per sample). In contrast, for the All-Voiced database, only one evaluation per rater was included to avoid potential biases from raters contributing multiple assessments, given the larger number of raters in this dataset. This methodological choice balanced the differences in dataset structures while preserving the integrity and richness of the PVQD data.

The evaluations from All-Voiced users were visually compared to the PVQD expert ratings, with voice competence levels distinguished by color. For each voice sample and attribute, variability across evaluators was illustrated. Cut-off values for overall severity, as established by Calaf and Garcia-Quintana (2024), were applied to facilitate the interpretation of the data. In addition to the visual analyses, descriptive statistics were calculated to further explore the evaluation data for voice samples, enabling the analysis of both standard perceptual attributes and less commonly evaluated features such as pitch, loudness, hypernasality, and fry.

A total of 264 participants took part in the study, with at least one new user registering each day since the app's launch. In terms of age distribution, the majority of participants were between 21 and 30 years old (49%), followed by 31–40 years (22%), 41–50 years (17%), and 51–60 years (7%). A smaller proportion of participants were under 21 years old (1%), 61–70 years (3%), and only 2 participants (1%) were between 71 and 80 years. There were no participants over 80 years old. With regard to gender pronouns, 88% of participants identified with “she/her,” while 8% identified with “he/him.” Additionally, 2% preferred “she/they,” 1% preferred “he/they”, and 1 user preferred other pronouns.

Looking at the country of residence, the majority of participants were from the United States (51%), followed by Spain (18%) and Ecuador (12%). Other countries with notable representation included Canada, Peru and France (7 users each, 3%), as well as the United Kingdom, and Australia (6 users each, 2%). Countries with fewer participants included Italy, Albania, and Belgium (2 users each, 1%) and a diverse range of countries with one participant each, including Estonia, Colombia, Greece, Pakistan, Brazil, Chile, Turkey, Argentina, Bolivia, Costa Rica, and Slovakia.

In terms of language preference, 57% of participants chose English, followed by Spanish (25%) and Catalan (6%). French was also notably represented (4%), while Italian and Portuguese were each selected by 2 participants (1%). Other languages, such as Persian, Turkish, Greek, Pashto, Tamil, and Slovak, were each selected by 1 participant.

Regarding profession, the majority of participants were speech-language pathologists (78%). Other professions included “other” (17%), singing teachers (1%), one otolaryngologist (ENT specialist), and one vocal coach. Finally, regarding voice competence level, 40% of participants classified themselves as beginners, 25% as intermediate, 20% as advanced, and 14% as experts.

A total of 557 evaluations were collected from users, reflecting a strong initial engagement with the platform. These evaluations encompassed 112 different voice samples, providing a diverse dataset for analysis. 60 evaluations, spanning 42 unique vocal samples, were voluntarily shared with the community, contributing to a collaborative feedback loop that enhances the training process for future participants. While the majority of samples received 1 or 2 evaluations (32 samples had 1 evaluation and 33 samples had 2 evaluations), certain voice samples were evaluated significantly more frequently, with some receiving evaluations from up to 23 unique raters.

Figs. 1–4 illustrate the distribution of Overall Severity, Roughness, Breathiness, and Strain scores for samples evaluated by 10 or more unique All-Voiced raters, compared with evaluations from the Perceptual Voice Quality Database (PVQD) by Walden (2022). In these graphs, the voice competence levels of the study participants are distinguished by different colors. This visual representation highlights how evaluators with different levels of voice competence assessed the same voice samples.

Complementing these visualizations, Table 1 presents descriptive statistics for the same attributes across three sources: All-Voiced raters, PVQD raters, and the combined dataset (All-Voiced+PVQD). This table offers detailed insights into central tendencies (mean, median) and variability (standard deviation [Std], interquartile range [IQR]) for each source.

As shown in Fig. 1 and detailed in Table 1, Overall Severity ratings reveal varying levels of agreement among raters across samples. High agreement is observed in samples such as PT012 (IQR=10), LA2007 (IQR=12), BL01, and NYU1028 (both IQR=14). Conversely, substantial dispersion is evident in NYU1029 (IQR=39) and NYU1030 (IQR=35). Notably, PT012 and LA2007 present smaller IQRs in the All-Voiced dataset (3 and 9, respectively) compared to the PVQD dataset (22 and 26). In contrast, samples like LA7008 and NYU1029 exhibit greater variability among All-Voiced raters (IQRs of 49 and 44) than PVQD raters (IQRs of 12 and 18).

As depicted in Fig. 2 and detailed in Table 1, samples with lower Roughness levels, such as LA2007, BL01, and PT003 (median=2 for all), exhibit the least variability, with IQRs of 4, 8, and 9, respectively. Conversely, PT012, which showed the highest agreement for Overall Severity (IQR=10), exhibits the greatest variability for Roughness (IQR=43). In this sample, agreement is higher among All-Voiced raters (IQR=34) compared to PVQD raters (IQR=49). Conversely, NYU1025 shows lower overall agreement (IQR=33), with PVQD raters exhibiting greater consensus (IQR=20) than All-Voiced raters (IQR=44).

As illustrated in Fig. 3 and summarized in Table 1, samples with lower levels of Breathiness (median=0), such as NYU1028 and BL01, exhibit the least variability, with IQRs of 1 and 3, respectively. Interestingly, PT012 displays the highest variability for Breathiness (IQR=56), consistent with its high variability in Roughness (IQR=43). However, this same sample displayed the highest agreement in Overall Severity (IQR=10). Comparisons between datasets reveal differing patterns: for LA1003, All-Voiced raters showed greater consensus, while for NYU1029, NYU1030, NYU1025, and PT003, PVQD raters exhibited higher agreement.

For Strain, samples with lower medians, such as BL01 (median=0, IQR=4) and NYU1028 (median=1, IQR=8), exhibit the least variability. In contrast, NYU1029, NYU1030, and NYU1025 show the greatest variability, with IQRs of 37, 45, and 56, respectively. Notably, NYU1029 had greater agreement among PVQD raters (IQR=20) compared to All-Voiced raters (IQR=40). Conversely, PT003, NYU1025, and LA2007 show higher agreement among All-Voiced raters (Fig. 4, Table 1).

Finally, outliers were observed across samples and attributes, originating from both datasets (Figs. 1–4). Some correspond to All-Voiced raters with varying levels of expertise (e.g., BL01, LA2007, LA7008, NYU1025), while others stem from PVQD raters (e.g., BL01, LA2007, PT012).

As shown in Table 2, twenty voice samples with at least three evaluations were identified for less commonly assessed features, including Pitch (Too High, Too Low), Loudness (Too Loud, Too Soft), resonance-related comments (e.g., Hypernasality), and additional features (e.g., Fry). The number of evaluations per sample ranged from 3 to 9, with most evaluated by 3 to 5 raters and a few by as many as 7 or 9.

Variability in the evaluation of these less commonly assessed features is considerable. For example, for “Pitch too low,” the IQR ranges from 3 to 43, revealing a significant disagreement among raters for some samples, “Loudness too soft” shows IQRs ranging from 0 to 36, while “Fry” exhibits IQRs ranging from 6 to 30, reflecting varying levels of dispersion across samples. With respect to “Hypernasality”, despite significance is limited, as it was evaluated in only 2 samples, its assessment shows a notable variability, with IQRs ranging from 5 to 26.

The variation in the number of evaluations per sample suggests that the app is being actively used in academic settings. While the majority of the 120 samples had received only 1 or 2 evaluations 75 days after the launch, certain voice samples were evaluated significantly more frequently, with some reaching up to 23 unique raters. This disparity indicates that instructors may have selected specific samples for use in structured classroom activities, leveraging the app as a teaching tool. This aligns with the app's dual purpose: supporting autonomous training through randomly selected samples and facilitating guided instruction by enabling educators to target specific voice characteristics. These findings underline the app's potential for enhancing both individual learning and collaborative classroom-based training.

The distribution of voice competence levels among participants further supports the idea that the app is being incorporated into educational environments. A notable 40% of participants identified themselves as beginners, followed by 25% as intermediate, 20% as advanced, and 14% as experts. The high proportion of beginner users aligns with the app's potential use in teaching settings, where students learn auditory-perceptual skills under the supervision of instructors.

Importantly, it should be noted that registering for the app is not required for classroom use, yet many beginner participants chose to register and provide consent for their data to be used in research. This suggests that these users found additional value in the app beyond their initial class activities, motivating them to engage more deeply with the platform.

The distribution of participants by country and language can be attributed to several factors. First, the majority of participants in this study were from the United States (51%), which can largely be explained by the fact that the vocal sample database used in the app (Walden, 2022) is in English, making it more accessible to English-speaking professionals. While the app has been developed to support multiple languages with different CAPE-V adaptations, the sentences in the vocal samples remain limited to English at this stage. Efforts are underway to gather vocal samples in additional languages or to source from global databases that meet ethical requirements, including informed consent. This stepwise approach ensures immediate availability and utility, while laying the groundwork for greater linguistic diversity in future updates.

Additionally, Spain accounted for 18% of participants, with a significant portion selecting Catalan as their preferred language (half of them). This distribution is likely influenced by the professional connections of the app's developer, who has a strong network in Spain, including many Catalan speakers, which likely contributed to early registrations from this region.

Beyond these main regions, the app is also attracting users from a diverse array of countries and languages. This diversity highlights the app's growing international reach and its potential to engage a broad spectrum of users across linguistic and cultural contexts, even as efforts continue to expand the linguistic diversity of the vocal sample database.

An advantage of the data collection process in All-Voiced lies in its ability to digitize the CAPE-V comprehensively, capturing not only the core attributes of overall severity, roughness, breathiness, and strain but also additional features such as distinctions in pitch (e.g., too high or too low) and loudness (e.g., too loud or too soft), resonance-related comments, and additional features. While these attributes are already part of the CAPE-V framework, they are often overlooked in research due to the absence of tools that facilitate their systematic collection. By enabling the efficient capture of all CAPE-V attributes, All-Voiced allows for the identification of voice samples that exemplify these characteristics, which can be utilized in training programs and further studies. This comprehensive approach represents a significant step forward in fully leveraging the CAPE-V's potential for evaluation and research.

To this respect, the results of pitch and loudness evaluations from the PVQD (Walden, 2022) have not been included because it is unclear whether their ratings for inadequate pitch or loudness refer to an alteration that is too high, too low, too loud, or too soft. Such ambiguity complicates the study of reliability in these perceptual judgments. However, by introducing the directional distinctions in the data collection process, the All-Voiced app offers the potential to gather robust data for future studies on the reliability of perceptual judgments also for pitch and loudness alterations, thus addressing a critical gap in the current body of research.

The preliminary results of this study provide further insight into the variability observed in less commonly assessed features, underscoring the complexity of these evaluations. For instance, while some samples showed high consensus among raters, such as for “Pitch too low” with IQRs as low as 3, others exhibited substantial dispersion, with IQRs up to 43 for the same feature. Similarly, “Loudness too soft” displayed variability ranging from 0 to 36, and “Fry” had IQRs from 6 to 30. This variation suggests that certain attributes are inherently more challenging for raters to assess consistently, emphasizing the need for continued research to better understand the factors contributing to this variability and to refine training approaches for these specific perceptual dimensions.

The analysis of Overall Severity ratings reveals patterns of agreement among raters that do not fully align with previous published reports, which suggested greater agreement when evaluating either normal voices or severely dysphonic voices compared to mild and moderate dysphonias (Gerratt et al., 1993; Kreiman & Gerratt, 2000; Kreiman et al., 1993). In this study, the samples BL01 and NYU1028, classified as normal based on the cutoff values proposed by Calaf and Garcia-Quintana (2024), exhibit high agreement (IQR=14), which is consistent with this tendency. However, LA2007 (classified as mild, IQR=12) and PT012 (classified as moderate, IQR=10) show an even greater agreement, presenting a notable incongruence with the literature. This discrepancy suggests that factors beyond overall severity, such as specific perceptual attributes, sample characteristics, or evaluator familiarity with certain vocal patterns, may influence agreement levels, underscoring the need for further investigation.

When analyzing specific perceptual attributes, the tendency for normal or less altered samples to achieve higher agreement among raters becomes more evident. For instance, Roughness evaluation in samples LA2007, BL01, and PT003 show the least variability (median=2, IQRs=4, 8, and 9, respectively), consistent with their closer alignment to non-altered voice qualities. Similarly, for Breathiness, NYU1028 and BL01 (median=0, IQRs=1 and 3, respectively) exhibit a high level of agreement. For Strain, BL01 (median=0, IQR=4) and NYU1028 (median=1, IQR=8) show a strong agreement, whereas LA2007 (median=2, IQR=35) presents greater variability, which may be attributed to its classification as mild in Overall Severity compared to BL01 and NYU1028, which were classified as normal. Further investigation is needed to fully understand the factors influencing variability, particularly in attributes like Strain, where agreement levels appear to fluctuate depending on the sample's overall severity classification.

Despite these insights, a key limitation of this study is the limited volume of data at this early stage, which restricts the scope of inter-rater and intra-rater reliability analyses. Many voice samples received only a single evaluation, making it impossible to assess reliability across evaluators for those samples. Moreover, the current dataset size does not yet allow for meaningful exploration of trends based on demographic factors such as age, gender, profession, or language. However, as data collection expands, these analyses will become feasible, offering deeper insights into the patterns of agreement and variability observed in this study.

The observed variability in evaluations across samples and attributes, even among expert evaluators, aligns with findings from prior studies on CAPE-V adaptations across different languages. While high inter-rater reliabilities have been reported for certain attributes like overall severity (Karnell et al., 2007; Mozzanica et al., 2014; Pommée, Mbagira, et al., 2024), attributes such as strain, pitch, and loudness consistently exhibit lower inter-rater reliability across multiple languages, including English (Zraick et al., 2011), Brazilian Portuguese (Behlau et al., 2022), and Tamil (Venkatraman et al., 2022). Similarly, intra-rater reliability has varied across studies, particularly for attributes like pitch and loudness (Behlau et al., 2022; Salary Majd et al., 2014).

Outliers observed across all levels of expertise, including expert evaluators, further underscore the inherent challenges in achieving consistency. Factors such as differences in internal standards and interpretations of perceptual attributes likely contribute to these discrepancies, reflecting the persistent difficulty of ensuring consensus among evaluators, even within the structured framework of the CAPE-V.

The results for PT012 exemplify how All-Voiced can identify voice samples that pose special challenges for evaluators. Despite displaying the highest agreement in Overall Severity ratings (IQR=10), this sample exhibited substantial variability in the evaluation of specific perceptual attributes, such as Breathiness (IQR=56), Roughness (IQR=43), and Strain (IQR=29). This discrepancy suggests differences in how evaluators interpret and apply these specific perceptual attributes, reflecting persistent challenges in achieving consensus. According to Nagle, Kempster, et al. (2024), clinicians have reported difficulties with attributes like strain, overall severity, and pitch in the context of the CAPE-V, indicating a need for clearer definitions and guidelines. To address this issue, the same authors have proposed the CAPE-Vr [Pre-print] (Kempster, Nagle, & Solomon, 2024), a revised version of the CAPE-V framework that aims to resolve these conceptual ambiguities and provide more precise guidance for clinical practice.

While standardizing definitions and procedures is essential, the findings presented here emphasize the need to address the inherently subjective nature of auditory-perceptual evaluations. To this respect, All-Voiced emerges as a useful tool to identify voice samples with poor expert consensus, providing a unique opportunity to investigate the root causes of these discrepancies. Such disagreements may stem from disparities in raters’ cognitive representations of voice attributes, or from specific acoustic features in the samples that influence raters’ perceptions differently. Future work will focus on implementing targeted training modules and validated auditory anchors, aiming to create a shared training platform that enhances the reliability of auditory-perceptual evaluations through consensus building and the adoption of universally shared criteria.