SongBsAb (NDSS 2025)

Audio Samples

Display Format:

input target singing voice                                                                                                                                                undefended output singing voice

input source singing voice                                                                                                                                               defended     output singing voice

Dual Prevention

Dataset: NUS-CMS-48 (English)        Target Singer A (Female) + Lyric a

Dataset: NUS-CMS-48 (English)        Target Singer B (Male) + Lyric b

Dataset: NUS-CMS-48 (English)        Target Singer C (Male) + Lyric c

Dataset: NUS-CMS-48 (English)        Target Singer D (Male) + Lyric d

Dataset: OpenSinger (Chinese)        Target Singer E (Male) + Lyric e

Dataset: OpenSinger (Chinese)        Target Singer F (Female) + Lyric f

Dataset: OpenSinger (Chinese)        Target Singer F (Female) + Lyric g

Transferability

Transferability of Identity Disruption

Target Singer G   Identity Encoder of Adversary: LSTM   Identity Encoder of Defender: XV

Target Singer G   Identity Encoder of Adversary: LSTM   Identity Encoder of Defender: Auto

Target Singer H   Identity Encoder of Adversary: LSTM   Identity Encoder of Defender: XV

Transferability of Lyric Disruption

Lyric h    Lyric Encoder of Adversary: Whisper-Medium   Lyric Encoder of Defender: Whisper-Tiny

Lyric i    Lyric Encoder of Adversary: Whisper-Medium   Lyric Encoder of Defender:  Wav2vec2

Lyric j    Lyric Encoder of Adversary: Whisper-Medium   Lyric Encoder of Defender:  Decoar2

Code

Our code is available at