Andrés Piñeiro-Martín

Mr. Andrés Piñeiro-Martín

Multimedia Technology Group, Universidade de Vigo, Vigo, Galicia, Spain
orcid

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
21 July 2023 in Electronics
Verified
Journal Article
Open access
PDF file
APAndrés Piñeiro-Martín
Andrés Piñeiro-Martín
MGMARIA CARMEN GARCIA-MATEO
MARIA CARMEN GARCIA-MATEO
LDLaura Docio-Fernandez
Laura Docio-Fernandez
MLMaría del Carmen López-Pérez
María del Carmen López-Pérez
Virtual assistants (VAs) have gained widespread popularity across a wide range of applications, and the integration of Large Language Models (LLMs), such as ChatGPT, has opened up new possibilities for developing even more sophisticated VAs. However, this integration poses new ethical issues and challenges that must be carefully considered, particularly as these systems are increasingly used in public services: transfer of personal data, decision-making transparency, potential biases, and privacy risks. This paper, an extension of the work presented at IberSPEECH 2022, analyzes the current regulatory framework for AI-based VAs in Europe and delves into ethical issues in depth, examining potential benefits and drawbacks of integrating LLMs with VAs. Based on the analysis, this paper argues that the development and use of VAs powered by LLMs should be guided by a set of ethical principles that prioritize transparency, fairness, and harm prevention. The paper presents specific guidelines for the ethical use and development of this technology, including recommendations for data privacy, bias mitigation, and user control. By implementing these guidelines, the potential benefits of VAs powered by LLMs can be fully realized while minimizing the risks of harm and ensuring that ethical considerations are at the forefront of the development process.

2 June 2023 in Preprints.org
Verified
Preprint
Open access
APAndrés Piñeiro-Martín
Andrés Piñeiro-Martín
CGCarmen Garci­a-Mateo
Carmen Garci­a-Mateo
LDLaura Docío-Fernández
Laura Docío-Fernández
MPMaría del Carmen López Pérez
María del Carmen López Pérez
Virtual assistants (VAs) have gained widespread popularity across a wide range of applications, and the integration of Large Language Models (LLMs) such as ChatGPT has opened up new possibilities for developing even more sophisticated VAs. However, this integration poses new ethical issues and challenges that must be carefully considered, particularly as these systems are increasingly used in public services: transfer of personal data, decision-making transparency, potential biases, and privacy risks. This paper, an extension of the work presented at IberSPEECH 2022, analyzes the current regulatory framework for AI-based VAs in Europe and delves into ethical issues in depth, examining potential benefits and drawbacks of integrating LLMs with VAs. Based on the analysis, this paper argues that the development and use of VAs powered by LLMs should be guided by a set of ethical principles that prioritize transparency, fairness, and harm prevention. The paper presents specific guidelines for the ethical use and development of this technology, including recommendations for data privacy, bias mitigation, and user control. By implementing these guidelines, the potential benefits of VAs powered by LLMs can be fully realized while minimizing the risks of harm and ensuring that ethical considerations are at the forefront of the development process.

24 January 2025 in Applied Sciences
Verified
Journal Article
Open access
PDF file
APAndrés Piñeiro-Martín
Andrés Piñeiro-Martín
FSFrancisco-Javier Santos-Criado
Francisco-Javier Santos-Criado
MGMARIA CARMEN GARCIA-MATEO
MARIA CARMEN GARCIA-MATEO
LDLaura Docio-Fernandez
Laura Docio-Fernandez
MLMaría del Carmen López-Pérez
María del Carmen López-Pérez
Large language models (LLMs) have revolutionized the field of artificial intelligence in both academia and industry, transforming how we communicate, search for information, and create content. However, these models face knowledge cutoffs and costly updates, driving a new ecosystem for LLM-based applications that leverage interaction techniques to extend capabilities and facilitate knowledge updates. As these models grow more complex, understanding their internal workings becomes increasingly challenging, posing significant issues for transparency, interpretability, and explainability. This paper proposes a novel approach to interpretability by shifting the focus to understanding the model’s functionality within specific contexts through interaction techniques. Rather than dissecting the LLM itself, we explore how contextual information and interaction techniques can elucidate the model’s thought processes. To this end, we introduce the Context-Driven Divergent Knowledge Evaluation (CDK-E) methodology, along with the Divergent Knowledge Dataset (DKD), for evaluating the interpretability of LLMs in context-specific scenarios that diverge from the model’s inherent knowledge. The empirical results demonstrate that advanced LLMs achieve high alignment with divergent contexts, validating our hypothesis that contextual information significantly enhances interpretability. Moreover, the strong correlation between LLM-based metrics and semantic metrics confirms the reliability of our evaluation framework.

3 February 2026 in IEEE Access
Verified
Journal Article
Open access
APAndrés Piñeiro-Martín
Andrés Piñeiro-Martín
MGMARIA CARMEN GARCIA-MATEO
MARIA CARMEN GARCIA-MATEO
LDLaura Docio-Fernandez
Laura Docio-Fernandez
MLMaría del Carmen López-Pérez
María del Carmen López-Pérez
This paper explores the robustness and generalisation capabilities of foundation speech models, such as Whisper, and Large Language Models (LLMs) with speech encoders, like Phi-4, when fine-tuned for the End-to-End (E2E) Spoken Language Understanding (SLU) task in low-resource languages. We investigate the impact of model scale, language-specific pretraining, and specialised encoding strategies for both Whisper-based and LLM-based SLU, while also exploring prompting techniques that enable LLMs to handle tasks such as intent classification and slot filling, bridging the gap between raw audio and language understanding. Our results demonstrate that a speech-conditioned LLM can perform instruction-driven SLU in a language entirely unseen during pretraining. By adapting only the speech encoder of Phi-4 and keeping the decoder frozen, the system achieves competitive performance in Galician through in-context supervision. In parallel, we showthat increasing model scale and applying language-specific ASR pretraining consistently boost performance across all SLU metrics, particularly under the challenging conditions defined in the FalAI dataset. Finally, we present a detailed ablation study, showing that while acoustic redundancy has limited impact beyond a certain threshold, lexical diversity plays a crucial role in supporting robust generalisation. These findings offer new insights into data efficiency and generalisation in low-resource settings.