Por favor, use este identificador para citar o enlazar este ítem: https://hdl.handle.net/20.500.12008/56847

Título:

Technical solutions for marking and detecting AI-generated audio content in the context of article 50(2) AI act : final study report

Otros títulos:

Coordinador:

Director:

Compilador:

Autor:

Serra, Xavier
Oguz Araz, Recep
Batlle-Roca, Roser
Juvela, Lauri
López-Ayala, David
Rocamora, Martín

Tutor:

Tipo de documento:

Reporte técnico

Editor:

Palabras clave:

Audio watermarking
AI detection
Trustworthy AI
EU AI Act

Descriptores:

Año de publicación:

2026

Contenido:

Resumen:

This report reviews the current state of technical solutions and research on marking and detecting AI-generated audio within the framework of Article 50 of the EU AI Act. It aims to support the development of Codes of Practice addressing transparency obligations by providing an informed and comprehensive analysis of existing methods, their capabilities, and limitations. Specifically, the report: (1) examines the principal mechanisms for audio marking and detection, (2) proposes a set of evaluation criteria aligned with the AI Act’s requirements, (3) assesses state of the art approaches against these criteria, (4) identifies their main shortcomings and technical gaps, and (5) outlines strategic directions for future research and standardization. The report distinguishes between marking, detection, and identification, which are often used interchangeably but serve distinct roles. Marking refers to the intentional embedding or attachment of information that signals AI generation (e.g., metadata or watermarks). Detection involves verifying whether such marks are present or have been altered. Identification extends further, aiming to infer the origin of the content — such as the generative model used — even when no explicit mark exists. This distinction underpins the comparative analysis and evaluation presented in the report. The analysis is based on an extensive review of technical literature, existing standards, and industrial practices, complemented by expert consultations involving researchers, industry specialists in watermarking and synthetic media detection, and policymakers. The evaluation shows that no single technology currently fulfils all requirements of Article 50, namely effectiveness, interoperability, robustness, and reliability. Metadata, particularly when combined with cryptographic protection, ensures transparency and ease of integration but is vulnerable to being stripped or altered. Audio watermarking enables the embedding of persistent identifiers but faces challenges in robustness and interoperability. Audio fingerprinting effectively verifies known content but cannot identify new or modified AI-generated material. Generative model identification can detect AI-specific artefacts without prior marking, yet it lacks generalization and is vulnerable to manipulation. Given these limitations, the report advocates for a multi-layered approach that combines complementary methods. A robust framework should integrate cryptographically protected metadata for transparency, imperceptible watermarking for persistence, and forensic detection techniques for unmarked content. Fingerprinting remains useful in specific, controlled scenarios where reference databases are available. Ensuring reliable deployment and regulatory compliance will require continued research and standardization, particularly to enhance robustness, interoperability, and adaptability to evolving generative models. Achieving these objectives calls for coordinated efforts among academia, industry, and regulators, supported by the European Commission’s role in promoting open standards, funding public infrastructures (e.g., reference datasets and benchmarks), and fostering collaborative ecosystems. Ultimately, this report contributes to the European Commission’s ongoing initiative to define Codes of Practice that guide the practical implementation of transparent, accountable, and technically sound solutions for the marking and detection of AI-generated audio.

Descripción:

Editorial:

Publications Office of the European Union

EN:

Technical Report, European Union, 2026, pp. 1-50, DOI: 10.2759/0784462.

Financiadores:

Comisión Europea

Citación:

Serra, X., Oguz Araz, R., Batlle-Roca, R. y otros. Technical solutions for marking and detecting AI-generated audio content in the context of article 50(2) AI act : final study report [en línea]. EN: Technical Report, European Union, 2026, pp. 1-50. DOI: 10.2759/0784462.

Citación:

metadata.cc.license.name:

Documento del Gobiberno:

URI:

https://hdl.handle.net/20.500.12008/56847

ISBN:

978-92-68-34421-7

e-ISBN:

ISSN:

ISMN:

Otros identificadores:

Título Obtenido:

Facultad o Servicio que otorga el Título:

Licencia:

Atribución/Reconocimiento 4.0 Internacional

Cobertura geográfica:

Unión Europea

Cobertura temporal:

metadata.dc.relation.IsVersionOf:

Clasificación:

Es versión de:

Colecciones:

Publicaciones académicas y científicas
FicherosDescripciónTamañoFormato
SOBJLR26.pdf Versión publicada 689.88 KB Adobe PDF
Mostrar el registro Dublin Core completo del ítem

Licencia Creative Commons:

Atribución/Reconocimiento 4.0 Internacional
Excepto donde se indique lo contrario, la licencia de este ítem se describe como Atribución/Reconocimiento 4.0 Internacional