SingFox: A Multi-Lingual Singfake Detection Corpus

Authors: Arth J. Shah, Devanshi K. Trivedi, Himanshi U. Borad, Hemant A. Patil

Published: 2026-06-17 12:07:20+00:00

Comment: Accepted at INTERSPEECH 2026

AI Summary

This work introduces SingFox, a comprehensive and large-scale multi-lingual dataset designed for evaluating singing deepfake detection and source tracing systems. The dataset features over 113,802 audio clips spanning 20 languages and 126.32 hours, divided into six tracks to assess model robustness against various forms of novelty, including language diversity, genre-specific music, and alternative fake generation methods. SingFox aims to provide a reliable benchmark for reproducibility and accelerated research in this emerging field.

Abstract

In this work, we introduce SingFox, a comprehensive and large-scale dataset specifically designed to support robust evaluation of singing deepfake detection and source tracing systems. SingFox is divided into six distinct tracks (T1--T6), each targeting a unique form of novelty, ranging from language diversity (global and Indian) to genre-specific music and alternative fake generation methods. The dataset encompasses over 113,802 audio clips across 20 languages, totaling more than 126.32 hours of audio data and featuring 1,150 singers. Each track is designed to emulate real-world scenarios and evaluate how reliably models perform under different conditions, thereby assessing their robustness. SingFox aims to foster reproducibility and accelerate research in singing deepfake detection by providing a reliable benchmark for both the singfake detection task and the source verification task (model explainability). Experimental results show a highest accuracy of 77.84\\% in cross-dataset evaluation settings. All code and resources required to reproduce the dataset are publicly available at https://github.com/Arth-Shah/SingFox.

Key findings

Experimental validation demonstrated that models trained on existing datasets often perform poorly on SingFox due to generalizability issues, with the highest cross-dataset accuracy of 77.84% achieved by a model trained on the FMC dataset. Alternative fakes (real music with fake vocals) presented a significant challenge, causing a sharp drop in accuracy to 45.13%. LFCC+ResNet showed robust performance in source tracing with 89.06% accuracy.

Approach

The paper's main contribution is the creation of SingFox, a novel dataset for singing deepfake detection. It systematically generates fake singing audios using diverse methods like GANs, Diffusion Models, Voice Conversion, and Text-to-Music, and structures the data into six tracks to evaluate different challenges such as language diversity, alternative fakes, and source tracing.

Datasets

SingFox (proposed), CtrSVDD, WildSVDD, FMC

Model(s)

LFCC+ResNet, MFCC+CNN/BiLSTM/BiGRU, GFCC+ResNet, SSL-based Wav2Vec2 features

Author countries

India

← Previous