Joint Fullband-Subband Modeling for High-Resolution SingFake Detection

Xuanjun Chen; Chia-Yu Hu; Sung-Feng Huang; Haibin Wu; Hung-yi Lee; Jyh-Shing Roger Jang

cs.SDApr 6, 2026

Joint Fullband-Subband Modeling for High-Resolution SingFake Detection

Authors: Xuanjun Chen, Chia-Yu Hu, Sung-Feng Huang, Haibin Wu, Hung-yi Lee, Jyh-Shing Roger Jang

Organizations: Graduate Institute of Communication Engineering, National Taiwan University · Department of Computer Science and Information Engineering, National Taiwan University · NVIDIA Taiwan · Independent Researcher · NTU Artificial Intelligence Center of Research Excellence, National Taiwan University

Abstract

Rapid advances in singing voice synthesis have increased unauthorized imitation risks, creating an urgent need for better Singing Voice Deepfake (SingFake) Detection, also known as SVDD. Unlike speech, singing contains complex pitch, wide dynamic range, and timbral variations. Conventional 16 kHz-sampled detectors prove inadequate, as they discard vital high-frequency information. This study presents the first systematic analysis of high-resolution (44.1 kHz sampling rate) audio for SVDD. We propose a joint fullband-subband modeling framework: the fullband captures global context, while subband-specific experts isolate fine-grained synthesis artifacts unevenly distributed across the spectrum. Experiments on the WildSVDD dataset demonstrate that high-frequency subbands provide essential complementary cues. Our framework significantly outperforms 16 kHz-sampled models, proving that high-resolution audio and strategic subband integration are critical for robust in-the-wild detection.

Figures & tables

Figure 1: Comparison of audio spectral coverage under different sampling rates. Existing systems typically process 16 kHz sampled audio, restricting them to the speech-critical band (0–8 kHz) and discarding high-frequency details. In contrast, our approach utilizes 44.1 kHz audio to cover the full spectral range (0–22.05 kHz). This preserves extended harmonics and breath textures essential for detecting sophisticated singing forgeries that are mathematically invisible at lower sampling rates.

Figure 2: The overview of our proposed Sing-HiResNet framework. The framework is implemented in two stages: Phase 1 establishes the backbone for fullband and subband expert models, while Phase 2 facilitates their integration through various joint fusion processes.

Fullband $N=1$
Model	Subband (kHz)	Valid	Test A	Test B
FB [0.0,22.0]	0.00–22.05	0.66	2.31	10.79
Subband $N=2$
SB [0.0,11.0]	0.00 – 11.03	1.39	2.81	22.01
SB [11.0,22.0]	11.03 – 22.05	1.71	5.66	15.32
SB-Concat-2		0.09	3.13	16.29

Table 1: EER (%) of subband expert models across different subband partition configurations. Model names indicate approximate frequency ranges for readability, while the subband (kHz) column reports the exact band boundaries. SB-Concat-

N

denotes the feature-level concatenation model combining all subbands in the same configuration.

	Selected Model Pool $\mathcal{P}$				Aggregation		Concatenation		Interaction		Distillation *
	FB	SB ${}_{\texttt{L}}$	SB ${}_{\texttt{M}}$	SB ${}_{\texttt{H}}$	Test A	Test B	Test A	Test B	Test A	Test B	Test A	Test B
Pool $\mathcal{P}_{a}$	✓	✓			1.73	10.79	2.95	15.53	2.99	15.32	1.83	11.00
Pool $\mathcal{P}_{b}$	✓			✓	2.56	11.21	2.95	20.82	3.24	16.29	3.13	12.19
Pool $\mathcal{P}_{c}$	✓		✓		2.28	10.79	2.85	16.50	3.02	15.32	2.17	10.79
Pool $\mathcal{P}_{d}$	✓	✓		✓	1.87	9.82	3.57	19.64	3.02	16.50	1.87	12.19
Pool $\mathcal{P}_{e}$	✓	✓	✓		1.73	8.84	3.43	24.16	3.75	7.45	1.65	9.06

Table 2: EER (%) of the proposed integration framework across different component configurations.

Figure 3: EER (%) results across two categorization schemes. The left columns present a method-centric view to highlight frequency impact, while the right columns provide a condition-centric comparison of integration strategies across Test A and Test B.

Figure 4: Grad-CAM visualizations of expert and distilled models for bonafide and deepfake samples, featuring (a) single-teacher (Low) and (b) dual-teacher (Low/Mid-High) distillation. White dashed lines marking corresponding subband boundaries.

	Notation	Description	Methods	Selected Model Pool $\mathcal{P}$				EER (%)
	Notation	Description	Methods	FB ${}_{\texttt{D-LM}}$	FB	SB ${}_{\texttt{L}}$	SB ${}_{\texttt{M}}$	Test A	Test B
(a)	FB ${}_{\texttt{D-LM}}$	Pool $\mathcal{P}_{e}$ in Table 2	Distillation *		✓	✓	✓	1.65	9.06
(b)	FB ${}_{\texttt{A-LM}}$	Pool $\mathcal{P}_{e}$ in Table 2	Aggregation		✓	✓	✓	1.73	8.84
(c)	FB ${}_{\texttt{SA-D-LM}}$	Subband	Aggregation	✓		✓	✓	1.73	12.19
(d)	FB ${}_{\texttt{FSA-D-LM}}$	Fullband-Subband	Aggregation	✓	✓	✓	✓	1.58	8.77
* FB is the fullband student model, SB ∼ are the subband expert teacher models.

Table 3: EER (%) of cross-integration strategies. SB

{}_{\texttt{L}}

and SB

{}_{\texttt{M}}

denote expert models for 0–11 kHz and 11–16.5 kHz bands, respectively.

Models	Backbone	$f_{s}$	Test A (CI)	Test B (CI)
Baseline1	Wav2vec	16	6.09	24.09
Baseline2	Raw	16	8.84	26.11
PDL	Ensemble	16	5.80	22.01
NTU	SingGraph	16	4.31	31.82
IMS-SCU	WavLM	16	3.54	15.32
IMS-SCU	Ensemble	16	2.70	12.95

Table 4: Performance comparison on WildSVDD dataset, where

f_{s}

denotes sampling rate (kHz). Results are presented in Equal Error Rate (EER, %) with 95% Confidence Intervals (CI).

Explore similar work

From Talking to Singing: A New Challenge for Audio-Visual Deepfake Detection
May 27, 2026Ke Liu, Jiwei Wei, Wenyu Zhang +5Audio Deepfake Detection Domain Generalization
Center for Future Media, School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China
Escaping the Linearity Trap: Manifold Detours for Black-Box Adversarial Attacks on Singing Audio Deepfake Detection
May 18, 2026Yifan Liao, Yule Liu, Zhen Sun +5Audio Deepfake Detection Transfer-Based Adversarial Attacks
Wuhan University · The Hong Kong University of Science and Technology (Guangzhou)
Time-Frequency Consistency Learning for Robust Speech Deepfake Detection
Jul 20, 2026Jun Xue, Zhuolin Yi, Yanzhen Ren +6Audio Deepfake Detection Frequency-Domain Feature Learning
School of Cyber Science and Engineering, Wuhan University Wuhan, China · Wuhan University Wuhan, China · Tongji University ShanghaiU, China

May 27, 2026J/K move · Enter open · S save

From Talking to Singing: A New Challenge for Audio-Visual Deepfake Detection

Ke Liu, Jiwei Wei, Wenyu Zhang, Shuchang Zhou+4

Center for Future Media, School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China

Audio Deepfake Detection Domain Generalization

Open paper AskPDF arXiv

With rapid advances in audio-visual generative models, reliable forgery detection becomes increasingly critical. Existing methods for audio-visual deepfake detection typically rely on cross-modal inconsistencies. In singing, rhythmic vocalization weakens this coupling and introduces a nontrivial domain shift, substantially degrading detection performance. We construct the Singing Head DeepFake (SHDF) dataset using rhythm-aware generative models to fill the gap in singing benchmarks. To cope with cross-scenario domain shifts, we propose a Text-guided Audio-Visual Forgery Detection (T-AVFD) framework that generalizes across both talking and singing scenarios. T-AVFD comprises a facial authenticity pattern learner and a multi-modal differential weight learning module. The pattern learner aligns facial features with multi-granularity textual descriptions to learn generalizable authenticity patterns. The weight learning module preserves intrinsic audio-visual consistency and adaptively integrates it with authenticity patterns via differential weighting. Extensive experiments on multiple talking head deepfake datasets and SHDF show consistent improvements over existing baselines and strong robustness under diverse perturbations.

Abstract

Explore similar work

From Talking to Singing: A New Challenge for Audio-Visual Deepfake Detection

Escaping the Linearity Trap: Manifold Detours for Black-Box Adversarial Attacks on Singing Audio Deepfake Detection

Time-Frequency Consistency Learning for Robust Speech Deepfake Detection