Joint Fullband-Subband Modeling for High-Resolution SingFake Detection

A research team including Xuanjun Chen, Hung-yi Lee, and Jyh-Shing Roger Jang has developed a novel AI system for detecting AI-generated singing voices, known as SingFake or Singing Voice Deepfake Detection (SVDD). The key innovation is the first systematic use of high-resolution 44.1 kHz audio, a standard for music, instead of the 16 kHz audio common in speech-focused detectors. This captures vital high-frequency information that conventional models discard, which is essential for analyzing singing's complex pitch, wide dynamic range, and timbral variations.

The team's proposed solution is a joint fullband-subband modeling framework. This architecture uses a fullband model to capture the global context of the audio, while separate subband-specific expert models isolate fine-grained synthesis artifacts that are unevenly distributed across the frequency spectrum. Experiments on the challenging WildSVDD dataset demonstrate that these high-frequency subbands provide essential complementary cues for detection. The framework significantly outperforms existing 16 kHz-sampled models, establishing that high-resolution audio and strategic subband analysis are critical for building robust detectors that can work 'in-the-wild' against advanced singing voice synthesis.

This research, submitted to INTERSPEECH 2026, addresses a growing security gap as AI voice synthesis technology rapidly advances, increasing the risks of unauthorized imitation and deepfake misuse in the music industry. By moving to a music-standard sampling rate and designing a model that can learn from both broad and narrow frequency features, the team has created a more formidable technical barrier against audio forgeries.