The rapid implementation of digital services in financial, governmental, and commercial domains is accompanied by a growing demand for reliable user identification tools. Voice-based biometric authentication systems represent one of the most promising directions, as they combine ease of use with the potential for integration across a wide range of services-from banking operations and call centers to voice assistants. At the same time, such systems face a number of challenges: the need to scale to tens of thousands of users, to ensure resilience against attacks employing synthetic speech, and to maintain high accuracy in real-time operation. Neural network architectures capable of generating robust embeddings and preserving stability as data volumes increase are of particular importance. This article examines the scalability challenges of modern voice biometrics models and substantiates approaches to enhancing their efficiency, including architectural optimization, indexed search, and the use of representative speech corpora. The study emphasizes the necessity of a comprehensive approach to the development of voice authentication systems, in which technical performance is combined with security requirements and protection against cyber threats.
The scalability of a biometric system is defined by its ability to maintain efficiency as the number
of users and the volume of data increase. Biometric technologies—such as fingerprint, face, voice,
and iris recognition—must ensure high accuracy and rapid response even when operating with
millions of enrolled templates [
The analysis of recent scientific sources demonstrates a significant growth of interest in scalable solutions in the field of voice biometrics. Current research focuses on the development of embedding models, architectural optimization, and the applied use of technologies in banking services, call centers, and voice assistants.
In [
A more comprehensive comparative study was presented by Brydinskyi et al. (2024) [
Continuing the topic of optimization, Thienpondt and Demuynck (2023) [
Another relevant challenge is speaker identification from short speech fragments. Deng et al.
(2025) [
A broader historical perspective on the evolution of technologies was provided by Sharma et al.
(2024) [
The issue of security in scalable systems was addressed by Chen et al. (2023) [
Another important security aspect is privacy and heterogeneous training conditions. Chen & Xu
(2023) [
Meanwhile, even classical architectures remain subject to optimization. Sharif-Noughabi et al.
(2025) [
In summary, the analysis of recent publications indicates that the scalability of voice biometric systems relies on compact and robust embeddings, lightweight architectures, quantization methods, and solutions for short speech recordings, complemented by security and privacy mechanisms. Modern neural models are moving toward unifying high accuracy, optimal inference speed, and resilience to attacks, thereby opening the way for large-scale deployment in banking services, call centers, and voice assistants.
Highly accurate neural networks are often very large, which complicates their deployment on
devices with limited resources or under large-scale cloud access. To address the computational
challenges posed by such models, optimization methods are applied, in particular quantization—
reducing the dimensionality of neural networks. For instance, Amazon reduced the Alexa voice
assistant model to less than 1% of its original size without significant loss of accuracy [
Another critical component of scalable biometric systems is training data. The quality, size, and representativeness of training corpora directly influence model performance and adaptability. To ensure reliable operation across diverse conditions, models must be trained on large and heterogeneous datasets. Representativeness is especially important: datasets should cover a wide variety of accents, languages, and user groups to avoid bias and accuracy drops for underrepresented populations. For example, voice services trained predominantly on a single demographic may perform poorly for other groups with distinct accents, timbres, or vocabulary. Thus, investment in the collection and annotation of balanced, diverse training data is fundamental for scalable and universal AI systems.
Scalability is also inseparable from reliability. In high-availability services, such as banking voice platforms or consumer voice assistants, models must operate continuously (24/7). This requires deployment in clustered environments with effective load balancing across nodes. Such architectures prevent bottlenecks, ensure stable performance under peak loads, and support automatic failover: if one node fails, others seamlessly take over the workload. This approach guarantees service continuity, fault tolerance, and minimizes downtime risks—critical for online services, financial platforms, healthcare applications, and other domains where even brief disruptions can result in financial losses or reduced user trust.
Accordingly, system scalability encompasses not only the ability to process increasing data volumes and user requests, but also the assurance of robustness, fault tolerance, and uninterrupted service availability under real-world conditions.
For the experimental study of neural model scalability in voice biometric authentication systems, a
multilingual audio dataset was constructed, including both Ukrainian and English speakers [
To evaluate model scalability under conditions of linguistic variability, the dataset was supplemented with recordings of 20 English-speaking individuals, including well-known actors, journalists, and public figures. These audio materials were manually collected from open sources, primarily video interviews published on YouTube.
To prevent overlap between training and testing examples, an extended dataset was prepared. Specifically, for each of the 70 speakers, an additional 20 unique audio recordings were created, which did not duplicate the samples used for constructing reference embeddings. This ensured the correctness of verification accuracy evaluation and the validity of experimental results.
In the first stage of the study, all audio files belonging to a single speaker were processed by each of the selected neural networks: Pyannote, WavLM base-sv, WavLM base-plus-sv, and ECAPA. As a result, 10 embeddings were generated and then averaged into a single vector—the speaker’s reference embedding. Averaging reduces the impact of natural voice variations (intonation, tempo, background noise). The resulting reference embeddings were stored in a database together with a unique speaker identifier.
In the next phase of the study, it was necessary to determine the optimal verification threshold.
To this end, a full authentication process was simulated to assess system behavior under real-world
conditions. The objective was to achieve a balanced trade-off between false negatives and false
positives. Within this simulation, pairs of recordings were formed to cover two scenarios:
verification of valid users, where test samples were compared with the reference embeddings of the
same individual; and verification of invalid users, where test embeddings were compared with the
reference vectors of other speakers [
Once the reference database was created, the system proceeded to the verification stage. Using the same neural model employed for reference construction, an extended set of voice samples from each speaker was processed to extract test embeddings. Each test vector was then compared with the corresponding reference embedding using cosine similarity. If the cosine distance value was below a predefined threshold, verification was considered successful; otherwise, the sample was not recognized as belonging to the claimed user (Figure 1).
To assess the results of the experimental study, a set of commonly accepted metrics traditionally used in biometric verification was applied. The most frequently considered indicators for quantitative analysis of system performance are Accuracy, False Acceptance Rate (FAR), False Rejection Rate (FRR), and the integral criterion Equal Error Rate (EER).
Accuracy is a basic metric that reflects the overall proportion of correctly classified cases in the binary speaker verification task. It is defined as the ratio of correct decisions (successful authentications and correct rejections) to the total number of verification attempts. Despite its intuitive clarity, this metric is sensitive to class imbalance and is therefore not considered decisive in verification studies.
False Acceptance Rate (FAR) characterizes the security of the system and represents the proportion of cases in which an unauthorized user is incorrectly accepted as legitimate. Formally, FAR is calculated as the ratio of false acceptances to the total number of access attempts made by impostors. Minimizing FAR is critically important for improving system resilience to attacks and ensuring reliability.
False Rejection Rate (FRR), on the contrary, reflects system usability and measures the proportion of cases in which a registered user is incorrectly rejected as unauthorized. FRR is calculated as the ratio of false rejections to the total number of access attempts made by genuine users. A low FRR ensures better user experience and higher system availability. Obviously, it is impossible to minimize both FAR and FRR simultaneously. The optimal balance between them is achieved by selecting an appropriate similarity threshold for embeddings, which determines the priority: minimizing false acceptances or false rejections.
Equal Error Rate (EER) serves as an integral indicator of biometric system quality, defined at the
point where FAR and FRR intersect. The lower the EER, the more effective the system is
considered. Due to its independence from a specific threshold value, this metric is widely used as a
universal criterion for comparing different biometric authentication algorithms [
A series of experiments was conducted to evaluate the scalability of the voice biometric authentication system at five levels: 10, 20, 35, 50, and 70 registered users. For each level, comparative testing was performed on four neural models—Pyannote, WavLM base-sv, WavLM base-plus-sv, and ECAPA—followed by analysis of accuracy, false acceptance rate (FAR), false rejection rate (FRR), and the equal error rate (EER).
Table 1 presents the EER results obtained from the scalability experiments for the selected models.
Model
Regardless of scale, ECAPA-TDNN demonstrates the best performance: the EER remains within approximately 1.2–3.0% and consistently achieves the lowest value at each stage. Pyannote ranks second (about 2.7–5.4%), showing a moderate increase in error with the growing number of users and a slight improvement at 70 speakers. Both versions of WavLM exhibit the highest EER values (approximately 7–11%). A general trend is observed (Figure 2): as the number of speakers increases, the EER rises; however, partial stabilization occurs at the largest dataset size.
Based on the constructed graph of accuracy versus the number of users, it was established that the ECAPA-TDNN model demonstrates the highest stability and accuracy across all configurations. Its accuracy remains in the range of 95–98%, and even under maximum scaling (70 users) it maintains a high level of performance. Pyannote also showed stable accuracy, only slightly behind ECAPA-TDNN, with a mostly uniform degradation as the number of users increased.
The WavLM base-sv and WavLM base-plus-sv models demonstrated somewhat lower accuracy. A particularly noticeable decline was observed after exceeding the threshold of 35 users. This may indicate that, for these models, the increasing number of speakers complicates the discrimination of embeddings, leading to a higher number of system errors.
However, an interesting phenomenon was observed in the final iteration (70 users): the accuracy of Pyannote, WavLM base-sv, and WavLM base-plus-sv increased compared to the previous stage, breaking the general trend of degradation. This can be attributed to the larger pool of speakers, which allowed the models to establish better threshold values and reduce both false acceptances and false rejections.
Overall, the observed trend confirms the classical principle: as the number of users in the system increases, the task of distinguishing between voice embeddings becomes more complex, which can lead to higher error rates and reduced accuracy.
A histogram was constructed based on three key metrics: False Acceptance Rate (FAR) (red), False Rejection Rate (FRR) (blue), and Equal Error Rate (EER) (green). This visualization enabled a deeper analysis of model behavior under scalability conditions. The results provide a comprehensive assessment of the scalability and robustness of different architectures in the speaker verification task.
ECAPA-TDNN confirmed its high effectiveness: it maintained minimal EER values across all scalability levels and demonstrated a balanced trade-off between FAR and FRR. This indicates optimal threshold calibration and the model’s ability to generate discriminative embeddings that remain robust as the user base grows. Such characteristics make ECAPA-TDNN the most suitable choice for real-world deployment in systems with strict requirements for performance and security.
Pyannote also showed relatively low error rates; however, at higher scales (50–70 users), a gradual increase in FRR was observed. This reflects a reduced ability of the model to correctly recognize legitimate users as the database expands. Despite this, the overall error levels remain acceptable, making Pyannote suitable for medium-scale applications.
In contrast, WavLM base-sv revealed significant limitations: elevated FAR values indicate a higher probability of false acceptances. Such behavior is critically dangerous in security-sensitive scenarios, such as banking services or government registries, where even isolated false acceptances can have severe consequences.
The WavLM base-plus-sv variant slightly reduced EER compared to the base version but did not demonstrate stability comparable to ECAPA or Pyannote. As the number of users increased, FRR rose substantially, indicating the model’s declining ability to reliably recognize legitimate speakers. This significantly reduces usability and can negatively affect user experience.
The overall trend shows that scaling the user base is not a neutral factor across all models: the most robust architectures (such as ECAPA) maintain stable or even improved EER values due to better alignment of embeddings, while weaker models exhibit simultaneous growth of both FAR and FRR.
Thus, the results confirm that the scalability of biometric systems directly depends on the choice of model architecture. Powerful solutions such as ECAPA provide high tolerance to user base expansion, while other architectures require additional measures—embedding optimization, adaptive threshold selection, or even retraining on representative data. This underscores the need not only for careful model selection but also for the application of comprehensive methods to maintain scalability in the development of real-world voice biometric systems.
This study presented an experimental comparison of four neural models for voice authentication in order to evaluate their scalability. It was established that ECAPA-TDNN demonstrates the highest stability and the lowest EER values across all user levels, confirming its suitability for large-scale real-time systems. The Pyannote model also maintains acceptable accuracy, although it is characterized by a gradual increase in FRR as the database expands. In contrast, the WavLM architectures revealed scalability limitations: the base-sv variant exhibited elevated FAR, while the base-plus-sv version showed instability in FRR.
The findings indicate that maintaining high system performance and reliability requires the use of compact embeddings, indexed search, and model optimization methods such as quantization and knowledge distillation. The practical significance of the results lies in their applicability for the development of scalable biometric services in domains such as banking, call centers, and voice assistants.
Declaration on Generative AI While preparing this work, the authors used the AI programs Grammarly Pro to correct text grammar and Strike Plagiarism to search for possible plagiarism. After using this tool, the authors reviewed and edited the content as needed and took full responsibility for the publication’s content.