Speaker Verification with Speech-Aware LLMs: Evaluation and Augmentation

Speaker Verification with Speech-Aware LLMs: Evaluation and Augmentation Thomas Thebaud ID 1,2,∗ , Yuzhe Wang ID 1 , Laureano Moro-Velazquez ID 1 , Jesus Villalba-Lopez ID 1,2 , Najim Dehak ID 1,2 1 Electrical and Computer Engineering Department, Johns Hopkins University, Baltimore, MD, USA 2 Human Language Technology Center of Excellence, Johns Hopkins University, Baltimore, MD, USA tthebau1@jhu.edu Abstract Speech-aware large language models (LLMs) can accept speech inputs, yet their training objectives largely emphasize linguis- tic content or specific fields such as emotions or the speaker’s gender, leaving it unclear whether they encode speaker iden- tity. First, we propose a model-agnostic scoring protocol that produces continuous verification scores for both API-only and open-weight models, using confidence scores or log-likelihood ratios from the Yes/No token probabilities. Using this protocol, we benchmark recent speech-aware LLMs and observe weak speaker discrimination (EERs above 20% on VoxCeleb1). Sec- ond, we introduce a lightweight augmentation that equips an LLM with ASV capability by injecting frozen ECAPA-TDNN speaker embeddings through a learned projection and train- ing only LoRA adapters. On TinyLLaMA-1.1B, the resulting ECAPA-LLM achieves 1.03% EER on VoxCeleb1-E, approach- ing a dedicated speaker verification system while preserving a natural-language interface. Index Terms: automatic speaker verification, speech-aware LLM, large language model 1. Introduction Large Language Models (LLMs) [1, 2, 3] have recently become essential components of modern artificial intelligence systems. They have evolved into versatile architectures capable of pro- cessing inputs and outputs from multiple modalities, such as images [1, 4] and speech [5, 6, 7]. Speech-aware LLMs, defined by their capacity to process audio inputs [8], extend traditional text-based transformers by incorporating acoustic front-ends or audio tokenizers, enabling them to process spoken language di- rectly rather than relying solely on automatic speech recogni- tion transcripts. This innovation opens the door to models that could reason jointly over linguistic content and paralinguistic cues. However, most speech-aware LLMs either focus on lin- guistic understanding [9, 10], or audio-based question answer- ing [11, 10, 12], with a few prosody-based closed-set classifica- tion tasks such as accent, gender and emotion recognition [10]. Automatic Speaker Verification (ASV) aims to determine whether two speech recordings originate from the same speaker. It is a critical component in numerous applications, including biometric authentication, but also personalized assistants and dialogue analysis. State-of-the-art ASV systems [13, 14] are usually based on speaker embeddings (x-vectors [15]), that en- code the identity of a speaker within a single vector. These systems are highly optimized for identity discrimination and achieve remarkable performance on benchmarks such as Vox- Celeb [16]. If they can be leveraged for a variety of purposes, ** indicates the corresponding author. from speech emotion recognition [17] to healthcare [18, 19] or speaker characterisation [20], they are generally narrow in scope: they do not reason over linguistic content, nor are they designed to integrate seamlessly with higher-level reasoning tasks. The emergence of speech-aware LLMs raises an intriguing question: can a single large-scale, general-purpose model also perform fine-grained biometric tasks such as speaker verifica- tion? More specifically, do these models internally encode suf- ficient speaker-discriminative information, and can this infor- mation be harnessed or enhanced through appropriate training strategies? If so, this would suggest a path toward unified ar- chitectures capable of both high-level reasoning and low-level acoustic discrimination, reducing the need for task-specific pipelines. In this work, we investigate the capabilities of current speech-aware LLMs for automatic speaker verification. We an- alyze whether their learned representations contain discrimina- tive information for speaker identity. Building on this analy- sis, we propose a fine-tuning strategy that augments two widely used open-weight small-scale LLMs, tinyLLaMA 1.1B [21] and Ministral3 3.3B [1], with speaker verification capabilities by utilizing a pre-trained ECAPA-TDNN network from the speech brain toolkit [22]. Our approach adapts these models to pro- duce speaker-discriminative representations while preserving their general modeling capacity. Our contributions are as follows: • We propose a model-agnostic protocol to evaluate speaker verification capability in speech-aware LLMs. • We show that off-the-shelf speech-aware LLMs exhibit weak speaker discrimination on VoxCeleb1, mainly relying on coarse speaker characteristics. • We introduce a lightweight augmentation that injects frozen ECAPA speaker embeddings with LoRA adaptation to equip LLMs with ASV capability, achieving near SOTA capabili- ties. 2. Related Works 2.1. Automatic Speaker Verification Automatic Speaker Verification (ASV) has been extensively studied over the past two decades, evolving from generative sta- tistical models to highly discriminative deep learning systems. Most of those systems have been evaluated on the Voxceleb1 dataset [23], a corpora of over a 100k utterances extracted from YouTube videos of 1251 celebrities, which contains 3 testing splits: Original, Extended and Hard (Vox1-O/E/H). The introduction of the i-vector [24] marked a major mile- arXiv:2603.10827v1 [cs.SD] 11 Mar 2026 stone, providing a compact representation of speaker char- acteristics that could be compared using probabilistic linear discriminant analysis (PLDA), which showed a 8.8% EER on Vox1-O trained under the same conditions [23]. The x- vector framework [15] replaced generative factor analysis with a deep neural network trained to classify speakers, extracting fixed-dimensional embeddings from variable-length utterances, which bought up to 44% relative improvement in EER in var- ious scenarios evaluated on SITW [25]. Subsequent work in- troduced architectural refinements and training objectives tai- lored to speaker discrimination. In particular, time-delay neu- ral networks (TDNNs) and their enhanced variants, such as ECAPA-TDNN [13], incorporate channel attention, multi-scale feature aggregation, and squeeze-and-excitation mechanisms to better capture long-range speaker characteristics, which pushed the EER down to 0.8% on Vox1-O. If more recent improve- ments have been produced since, we selected ECAPA-TDNN for our experiments for its accessibility through the speechbrain toolkit [22], maximizing the reproducibility of our results. 2.2. Speech-Aware Large Language Models Large Language Models (LLMs) are transformer-based archi- tectures trained on large-scale text corpora using self-supervised next-token prediction. This scaling has led to strong reasoning abilities, making LLMs a core component in many frameworks. Open-weight models such as LLaMA [2] and Ministral3 [1] provide efficient decoder-only architectures with competitive generalization, while proprietary systems such as GPT [3] and Gemini [26] extend these capabilities to multimodal reason- ing. Although originally designed for text, transformers are modality-agnostic and can process other inputs once mapped into a compatible embedding space. For speech, raw waveforms are typically encoded using pre- trained acoustic models such as HuBERT [27], WavLM [28], or Whisper [29], or discretized using speech tokenizers and neural codecs [30, 31, 32]. Speech-aware LLMs integrate these rep- resentations with pretrained language models through learned projection layers or adapters [33, 34, 5, 6, 35], enabling tasks such as speech recognition, spoken QA, and multimodal dia- logue [8]. However, their training objectives primarily target linguistic and semantic understanding, leaving open the ques- tion of whether they encode sufficiently discriminative speaker information for biometric tasks such as speaker verification. We will investigate some of those models, utilizing a range of ar- chitectures, such as Qwen-2.5-7B [6] and AudioFlamingo3 [36] which leverages a Whisper encoder for its audio inputs or Kimi- audio-7B [35] which also adds a neural audio codec for noise and music processing. 3. Methods 3.1. Dataset VoxCeleb1 [23] and VoxCeleb2 [37] are large-scale, pub- licly available audiovisual datasets widely used for speaker recognition research. VoxCeleb1 contains over 100,000 utter- ances from 1,251 celebrities, extracted from YouTube interview videos. VoxCeleb2 significantly expands this effort, comprising over one million utterances from 5,994 speakers. When training is needed, the experiment performed in this article will use the development set of VoxCeleb2 as training, and the testing set of VoxCeleb2 as validation. Both sets are defined in the original article and contain disjoint speakers. For ablation studies, we also define a smaller set of VoxCeleb2-dev, using only 10% of the speakers selected at random, and keeping only 10 utterances per speaker, which we name VoxCeleb2-dev-XS, containing 600 speakers, for 6k utterances and a total of 12.4h of audio. The testing sets used for all experiments will be the 3 test splits from VoxCeleb1 mentioned previously: the Original, Extended, and Hard trials, which are defined as lists of pairs of enroll and test utterances to compare. 3.2. Proposed Speaker-Aware LLM architecture Cascaded speech-aware LLMs are usually built around a pre- trained speech encoder, a connector, and a pretrained LLM, which are finetuned to work jointly. We propose a cascaded speaker-aware LLM, as shown in Figure 1, which is built using a pretrained ASV system, a connector, and a pretrained LLM: • The ASV system is an ECAPA-TDNN [13], trained on VoxCeleb2-dev using speechbrain toolkit [22], which shows 0.89%, 0.45% and 0.96% EER respectively on VoxCeleb1 Original, Extended and Hard splits using cosine scoring. This system will be frozen during training. • The connector is a linear layer, which is used to project the x- vectors in the dimension of the text embeddings of the LLM. • The LLM is either a TinyLLaMA 1.1B [21] or a Ministral3 3.3B [1]. Each model is finetuned using Lora adaptors [38]. The two Speaker-Aware configurations will be referred as SA- TinyLLaMA and SA-Ministral3 depending on the LLM used. 3.3. Metrics ASV is a binary classification task, where a pair of utterances is either from the same speaker (target) or not (non-target). Stan- dard evaluation of an ASV system relies on the computation of a likelihood score for a predefined list of target and non-target pairs of utterances, which is used to compute the Equal Error Rate (EER). This metric requires a continuous set of scores for evalua- tion. If open-source models allow access to intermediate repre- sentations and the likelihood of an answer for different tokens, closed-source models usually only output the text answer to a query, which makes the computation of a continuous score a harder task. The next section details the proposed techniques to obtain a score from speech-aware LLMs 3.4. Scoring speaker awareness in Speech-aware LLMs 3.4.1. Confidence scoring based on speaker characteristics In the case of closed-weight LLMs, or when the logits are not available, the only available output is the text. We propose to ask for a confidence score between 0 and 100 in the probabil- ity that two given utterances are from the same speaker. This confidence score can then be used to compute the EER. The proposed prompt is: These are two distinct audios. First, think about the elements that characterize each speaker, such as their gender, accent, tone, prosody, and speech rate. Give the characteristics for each audio Then, from those characteristics, infer the likelihood that both speakers are the same. Answer by Yes or No, and give a confidence score between 0 and 100: 0 corresponds to the certainty that they are from different speakers, 100 corresponds to the certainty that they are from the same speaker, And 50 means you are uncertain. First audio: [audio1] and second audio: [audio2]. Figure 1: Schematic of the pipeline to train and test a speaker-aware LLM. 3.4.2. Log-likelihood scoring based on logits For more open-weights LLMs, were the logits associated to an answer are accessible, we can simplify the prob- lem by asking a binary output, and compute the ratio of the logits for each answer. The simplified prompt would be: These are two distinct audios. First, think about the elements that characterize each speaker, such as their gender, accent, tone, prosody, and speech rate. Give the characteristics for each audio Then, from those characteristics, infer the likelihood that both speakers are the same. Answer by Yes or No. First audio:[audio1] and second audio:[audio2]. Then, we can access the logits probability for the token Yes and the logits for the token No, and compute a log likelihood ratio as: LLR(prompt) = log( p(Yes|prompt) p(No|prompt) )(1) This LLR can be used subsequently as a score to compute the EER, offering a much finer-grained analysis than the integer confidence score previously used. This technique is used to evaluate the capabilities of our two proposed cascaded speaker- aware LLMs: SA-TinyLLaMA and SA-Ministral3. 3.5. Experiments 3.5.1. Off-the-shelf speech-aware LLMs: confidence scoring We probe a set of off-the-shelf speech-aware LLMs using the confidence-scoring protocol described in Section 3.4.1. For each VoxCeleb1 trial (enrollment utterance, test utterance) we issue a single prompt containing the two audio segments and request a binary decision (same/different speaker) together with a confidence score in [0, 100]. The returned confidence is used as the trial score to compute EER on the VoxCeleb1 trial lists (Vox1-O/E/H). This protocol is used to evaluate GPT-4.0-audio [3] (model gpt-4o-audio-preview-2025-06-03), Qwen- 2.5-7B [6],Gemini 3-flash and 2.5-flash-lite[26],Au- dioFlamingo3 [36], and Kimi-audio-7B [35]. Failure rate: In addition to verification performance, we track the fraction of trials for which the model response cannot be parsed into a valid decision and confidence score (reported as failure rate). 3.5.2. Unprompted speaker characterisation As the prompt requests the model to consider paralinguistic el- ements such as the accent, gender, and prosody of the speaker, most models explicitly detail their perception of those qualities. We probe the outputs to measure the percentage of outputs con- taining gender and accents from the speakers, and report both the frequency of reporting and its accuracy. Gender Metrics: To simplify the evaluation process, only the outputs ’male’ or ’female’ are considered as reported gen- der 1 . The gender accuracy is computed using the gender labels provided in the metadata. According to the metadata, speakers in VoxCeleb1 are 44.84% female and 55.15% male. Accent Metrics: Accent accuracy is computed using the na- tionality labels provided in the metadata.Attribution of a speaker to any geographical location is considered as a success- ful accent prediction. If the model predicts a more restrictive label than the nationality (’London accent’ or ’Scottish accent’ for ’UK’ for example), the accent is counted right. A less restrictive accent (’Hispanic accent’ for ’Mexico’ for example) is counted as wrong. According to the metadata, speakers in VoxCeleb1 are 63.86% from the USA, 17.18% from the UK, 4.31% from Canada, and other 33 nationalities are each less than 3%. 3.5.3. Speaker-aware LLMs training and evaluation Training: All the proposed speaker-aware models are trained for 50 epochs on VoxCeleb2-dev, using a batch size of 64, a learning rate of 10 −4 for all trainable parts, using VoxCeleb2- test as a validation set, on a single Nvidia A100 80Gb GPU. The models are trained for next token prediction, to predict either ’Yes’ or ’No’, using batches composed of half target pairs and half non-target pairs. The model with the best validation EER across epochs is kept. The training and testing code for those models is available at this address 2 . Evaluation: We evaluate our speaker-augmented models using the log-likelihood scoring protocol from Section 3.4.2, where the verification score is the log-likelihood ratio (LLR) between the Yes and No tokens, and present the results for the 3 evalu- ation split of VoxCeleb1. Ablations: We train a variant where the LLM backbone re- mains frozen (no LoRA adapters) and only the connector is learned. We denote this model SA-TinyLLaMA F . This setting follows the frozen-LLM adaptation proposed in [33], which re- ported 12.08% EER on VoxCeleb1-O using LLaMA2 3B. Considering this problem could be seen merely as an align- ment problem between the speaker embedding space and the LLM embedding space, we hypothesize that only a fraction of the speakers would be enough. To that end, we train a model using the VoxCeleb2-dev-XS subset defined in section 3.1, and name this model SA-TinyLLaMA XS . To complete the abla- tion, a model is trained on the VoxCeleb2-dev-XS subset with a frozen TinyLLaMA, named SA-TinyLLaMA F XS . 1 No other expression of gender was noted in the outputs of any model. 2 github link will be available upon publication. Table 1: Speaker-awareness evaluation of a set of Speech-aware LLMs using confidence scoring. EER (%) is shown for each split of Voxceleb1-test, as well as the confidence score failure rate, accuracy and failure rate for gender and accent prediction. ModelEER↓FailureGenderGenderAccentAccent Vox1-OVox1-EVox1-HRate↓Accuracy↑Predicted↑Accuracy↑Predicted↑ Qwen-2.5-7B [6]37.01%34.83%45.43%0.68%97.98%99.82%75.45%76.76% Kimi-audio-7B [35]43.58%43.02%43.93%16.12%-0%-0% Gemini3-flash [26]45.13%45.64%44.26%0.38%92.16%99.75%84.99%85.72% Gemini2.5-flash-lite [26]36.15%37.09%47.55%16.56%91.40%99.91%83.32%83.90% GPT4.0-audio [3]22.62%21.88%38.91%0.05%97.32%99.42%82.65%82.90% AudioFlamingo3 [36]32.90%31.00%31.51%76.23%77.29%55.68%59.06%47.95% Table 2: EER results on the splits of VoxCeleb1 for our proposed speaker-augmented LLMs, evaluated using log- likelihood based scoring. The ECAPA-TDNN system is shown as a comparison, evaluated using cosine scoring. ModelVox1-OVox1-EVox1-H ECAPA-TDNN[13]0.89%0.45%0.96% SA-Ministral314.76%15.88%21.04% SA-TinyLLaMA1.87%1.03%2.20% SA-TinyLLaMA F 5.48%4.21%6.60% SA-TinyLLaMA XS 3.57%2.21%3.44% SA-TinyLLaMA F XS 27.01%27.82%28.55% 4. Results 4.1. Off-the-shelf speech-aware LLMs Table 1 shows that off-the-shelf speech-aware LLMs exhibit weak speaker discrimination under the confidence-scoring pro- tocol. EERs remain far above a dedicated ASV system, rang- ing from 22.62% (GPT-4.0-audio on Vox1-O) to approximately 45% (Gemini), with several models operating close to chance level (50% EER). These results indicate that, under standard instruction-following prompts, current speech-aware LLMs do not reliably expose speaker-discriminative information suffi- cient for verification, even when explicitly asked to reason about paralinguistic cues. Those poor performances could be ex- plained by the coarse granularity of the scores, effectively al- lowing for only 101 levels, but those are not even used by the models, as all the prompted models yielded between 12 and 16 different scores, mostly multiples of 10 and 5. We also observe substantial differences in robustness across models. In particu- lar, AudioFlamingo3 exhibits a high failure rate (76.23%), mak- ing its raw EER difficult to interpret. 4.1.1. Speaker Characteristics prediction Although verification performance is poor, several models are nevertheless able to infer coarse speaker attributes. For ex- ample, Qwen-2.5-7B, Gemini, and GPT-4.0-audio achieve high gender classification accuracy when a gender label is produced (92-98% in Table 1), with near-complete gender coverage. This contrast reveals a clear limitation: although the models cap- ture coarse speaker attributes, such information is insufficient for fine-grained speaker identity discrimination. This is partic- ularly evident on Vox1-H, where the absence of cross-gender trials reduces the usefulness of gender cues and leads to a no- ticeable performance drop. These findings support the hypoth- esis that current speech-aware LLM training objectives priori- tize linguistic and high-level paralinguistic features rather than identity-specific representations. 4.2. Speaker-aware LLMs results In contrast to off-the-shelf speech-aware LLMs, Table 2 shows that injecting frozen ECAPA speaker embeddings and train- ing only a lightweight connector together with LoRA adapters yields a large improvement in verification performance. SA- TinyLLaMA achieves EERs close to the ECAPA-TDNN co- sine baseline, indicating that a general-purpose LLM can be endowed with strong ASV capability when provided with an explicit speaker representation and minimal task-specific adap- tation, while preserving a natural-language interface. Ablation results further highlight the importance of adapt- ing the LLM backbone. When the LLM is frozen and only the connector is learned, performance degrades substantially: SA-TinyLLaMA F reaches 5.48% EER on Vox1-O. This con- firms that the gains are not solely attributable to the quality of the injected speaker embeddings; parameter-efficient adap- tation is necessary for the LLM to reliably interpret the speaker representation and produce stable verification decisions un- der the prompted output format. Finally, we observe that the smaller TinyLLaMA-1.1B backbone outperforms larger back- bones (LLaMA-3B and Ministral3-3B) in our current training setup, an effect that merits further investigation. The poor per- formances of Ministral3-3B is yet to be explained, but could be linked to the differences in the embedding space of the model, which may need a different losses for optimized training. 5. Conclusion This work investigated whether modern speech-aware LLMs encode speaker identity information and whether this can be leveraged for automatic speaker verification (ASV). We in- troduced a model-agnostic evaluation protocol that derives continuous verification scores from both API-only systems (via prompted confidence) and open-weight models (via log- likelihood ratios over Yes/No tokens). Benchmarking on Vox- Celeb1 shows that off-the-shelf speech-aware LLMs exhibit weak speaker discrimination, with EERs typically above 20%. We proposed a lightweight augmentation that injects frozen ECAPA-TDNN speaker embeddings into an LLM through a learned projection and LoRA adaptation. On TinyLLaMA- 1.1B, the resulting model approaches the performance of a dedicated ECAPA-TDNN system while preserving a natural- language interface. Overall, our results suggest a practical path toward speaker- aware LLMs by explicitly integrating strong speaker represen- tations rather than relying on implicit learning. Nevertheless, our evaluation is limited by the coarse and model-dependent nature of confidence-based scoring for closed systems, as well as by parsing failures in some APIs, which constrain direct comparisons. Future work will explore more robust scoring strategies and extend this framework toward tem- porally resolved speaker modeling, enabling tasks such as di- arization and multi-talker conversation analysis within speech- aware LLMs. 6. Acknowledgments 7. Generative AI Use Disclosure Parts of this manuscript were edited and rewritten for clarity, grammar, and style with the assistance of a generative AI lan- guage model. The model was used only for language refinement and formatting suggestions; it did not contribute to the scientific content of this work. In particular, no ideas, hypotheses, exper- imental designs, implementations, code, datasets, analyses, re- sults, or bibliographic references were generated by any GenAI system. All technical decisions and the final text were reviewed and verified by the authors, who take full responsibility for the content. 8. References [1] A. H. Liu, K. Khandelwal, S. Subramanian, V. Jouault, A. Ras- togi, A. Sad ́ e, A. Jeffares, A. Jiang, A. Cahill, A. Gavaudan et al., “Ministral 3,” arXiv preprint arXiv:2601.08584, 2026. [2] H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale et al., “Llama 2: Open foundation and fine-tuned chat models,” arXiv preprint arXiv:2307.09288, 2023. [3] OpenAI, J. Achiam, S. Adler, S. Agarwal, L. Ahmad et al., “Gpt-4 technical report,” 2024. [Online]. Available: https://arxiv.org/abs/2303.08774 [4] F. Shu, L. Zhang, H. Jiang, and C. Xie, “Audio-visual llm for video understanding,” arXiv preprint arXiv:2312.06720, 2023. [5] C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, and C. Zhang, “Salmonn: Towards generic hearing abilities for large language models,” arXiv preprint arXiv:2310.13289, 2023. [6] Y. Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,” arXiv preprint arXiv:2311.07919, 2023. [7] D. Li, C. Tang, and H. Liu, “Audio-llm: Activating the capabil- ities of large language models to comprehend audio data,” in In- ternational Symposium on Neural Networks. Springer, 2024, p. 133–142. [8] S. Arora, K.-W. Chang, C.-M. Chien, Y. Peng, H. Wu, Y. Adi, E. Dupoux, H.-Y. Lee, K. Livescu, and S. Watanabe, “On the landscape of spoken language models: A comprehensive survey,” arXiv preprint arXiv:2504.08528, 2025. [9] S. Shon, A. Pasad, F. Wu, P. Brusco, Y. Artzi, K. Livescu, and K. J. Han, “Slue: New benchmark tasks for spoken language un- derstanding evaluation on natural speech,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, p. 7927–7931. [10] B. Wang, X. Zou, G. Lin, S. Sun, Z. Liu, W. Zhang, Z. Liu, A. Aw, and N. Chen, “Audiobench: A universal benchmark for audio large language models,” in Proceedings of the 2025 Con- ference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2025, p. 4297–4316. [11] J. Lin, C. Zhang, T. Wang, and H. Li, “Audiorag: A challenging benchmark for audio reasoning and information retrieval,” arXiv preprint arXiv:2602.10656, 2026. [12] Z. Ma, Y. Ma, Y. Zhu, C. Yang, Y.-W. Chao, R. Xu, W. Chen, Y. Chen, Z. Chen, J. Cong et al., “Mmar: A challenging bench- mark for deep reasoning in speech, audio, music, and their mix,” arXiv preprint arXiv:2505.13032, 2025. [13] B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa- tdnn:Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,” arXiv preprint arXiv:2005.07143, 2020. [14] Y. Tu, W. Lin, and M.-W. Mak, “A survey on text-dependent and text-independent speaker verification,” IEEE Access, vol. 10, p. 99 038–99 049, 2022. [15] D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudan- pur, “X-vectors: Robust dnn embeddings for speaker recognition,” in 2018 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, p. 5329–5333. [16] K. Hechmi, T. N. Trong, V. Hautam ̈ aki, and T. Kinnunen, “Voxceleb enrichment for age and gender recognition,” in 2021 IEEE Automatic Speech Recognition and Understanding Work- shop (ASRU). IEEE, 2021, p. 687–693. [17] R. Pappagari, T. Wang, J. Villalba, N. Chen, and N. Dehak, “x- vectors meet emotions: A study on dependencies between emo- tion and speaker recognition,” in ICASSP 2020-2020 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, p. 7169–7173. [18] A. Favaro, L. Moro-Vel ́ azquez, A. Butala, C. Motley, T. Cao, R. D. Stevens, J. Villalba, and N. Dehak, “Multilingual evalua- tion of interpretable biomarkers to represent language and speech patterns in parkinson’s disease,” Frontiers in Neurology, vol. 14, p. 1142642, 2023. [19] K. L. Hom, H. Beigi, and R. Betti, “Application of speaker recog- nition x-vectors to structural health monitoring,” in Model Vali- dation and Uncertainty Quantification, Volume 3: Proceedings of the 39th IMAC, A Conference and Exposition on Structural Dy- namics 2021. Springer, 2021, p. 139–148. [20] D. Kwa ́ sny, P. Jemioło, and D. Hemmerling, “Explaining predic- tions of the x-vector speaker age and gender classifier,” in In- ternational Conference on Dependability and Complex Systems. Springer, 2021, p. 234–243. [21] P. Zhang, G. Zeng, T. Wang, and W. Lu, “Tinyllama: An open- source small language model,” 2024. [22] M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell, L. Lugosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong, J.-C. Chou, S.-L. Yeh, S.-W. Fu, C.-F. Liao, E. Rastorgueva, F. Grondin, W. Aris, H. Na, Y. Gao, R. D. Mori, and Y. Ben- gio, “SpeechBrain: A general-purpose speech toolkit,” 2021, arXiv:2106.04624. [23] A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “Voxceleb: Large-scale speaker verification in the wild,” Computer Speech & Language, vol. 60, p. 101027, 2020. [24] N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front-end factor analysis for speaker verification,” IEEE Trans- actions on Audio, Speech, and Language Processing, vol. 19, no. 4, p. 788–798, 2010. [25] M. McLaren, L. Ferrer, D. Castan, and A. Lawson, “The speakers in the wild (sitw) speaker recognition database.” in Interspeech, 2016, p. 818–822. [26] G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican et al., “Gemini: a family of highly capable multimodal models,” arXiv preprint arXiv:2312.11805, 2023. [27] W.-N. Hsu, B. Bolte, Y.-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,” IEEE/ACM transactions on audio, speech, and language processing, vol. 29, p. 3451–3460, 2021. [28] S. Chen, C. Wang, Z. Chen, Y. Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao et al., “Wavlm: Large-scale self- supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, p. 1505–1518, 2022. [29] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” in International conference on machine learning. PMLR, 2023, p. 28 492–28 518. [30] X. Zhang, D. Zhang, S. Li, Y. Zhou, and X. Qiu, “Speechtok- enizer: Unified speech tokenizer for speech large language mod- els,” arXiv preprint arXiv:2308.16692, 2023. [31] R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Ku- mar, “High-fidelity audio compression with improved rvqgan,” NeurIPS, vol. 36, p. 27 980–27 993, 2023. [32] A. D ́ efossez, J. Copet, G. Synnaeve, and Y. Adi, “High fidelity neural audio compression,” arXiv preprint arXiv:2210.13438, 2022. [33] T. Thebaud, Y.-J. Lu, M. Wiesner, P. Viechnicki, and N. De- hak, “Enhancing dialogue annotation with speaker characteristics leveraging a frozen llm,” arXiv preprint arXiv:2508.04795, 2025. [34] Q. Fang, S. Guo, Y. Zhou, Z. Ma, S. Zhang, and Y. Feng, “Llama- omni: Seamless speech interaction with large language models,” arXiv preprint arXiv:2409.06666, 2024. [35] D. Ding, Z. Ju, Y. Leng, S. Liu, T. Liu, Z. Shang, K. Shen, W. Song, X. Tan, H. Tang et al., “Kimi-audio technical report,” arXiv preprint arXiv:2504.18425, 2025. [36] A. Goel, S. Ghosh, J. Kim, S. Kumar, Z. Kong, S.-g. Lee, C.- H. H. Yang, R. Duraiswami, D. Manocha, R. Valle et al., “Audio flamingo 3: Advancing audio intelligence with fully open large audio language models,” arXiv preprint arXiv:2507.08128, 2025. [37] J. S. Chung, A. Nagrani, and A. Zisserman, “Voxceleb2: Deep speaker recognition,” arXiv preprint arXiv:1806.05622, 2018. [38] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large lan- guage models,” arXiv preprint arXiv:2106.09685, 2021.