Multilingual Deepfake Speech Dataset for Robust and Generalizable Detection

The rise of sophisticated technologies capable of generating realistic synthetic human speech has introduced significant security challenges in voice-based applications. These advances in speech generation have enabled malicious actors to produce convincing speech deepfakes, undermining the reliability of speaker verification systems and digital communication. Despite growing interest in deepfake speech detection, many existing datasets are limited in linguistic diversity and fail to capture the complexity of real-world scenarios, thus constraining model generalization. In this work, we introduce the JAIST Multilingual Deepfake Speech (JMDS) dataset, a large-scale, multilingual, and multi-source corpus designed to support the development and evaluation of robust and generalizable deepfake speech detection systems. Covering 17 languages and comprising over 350,000 utterances, JMDS incorporates a wide range of deepfake generation methods and includes both human (pristine) and machine-generated (spoofed) speech, sourced from publicly available corpora and a curated subset of private data. We provide detailed analyses of utterance duration, generation techniques, and audio quality, along with comprehensive evaluations across multiple model architectures and configurations. Cross-dataset evaluations are also conducted to assess the generalization capabilities of detection models across diverse languages and data domains. This study contributes to a deeper understanding of the limitations and opportunities in current detection systems, ultimately paving the way for more resilient and linguistically inclusive countermeasures.