- Download OpenSinger dataset from OpenSinger and then
tar -zxvf OpenSinger.tar.gz- Change the working directory.
cd Lora-SVC-
Build the environment by following the instructions at
setup.sh -
Process the dataset.
python preprocess_dataset_opensinger.py- Obtain the undefended output singing voices.
python few_shot_svc.py -start 0 -end 76 --type all_singers-10_voices --source_type '' -dataset opensinger -w_start 0 -w_end 10000- Generate adversarial examples for input target singing voices.
python protect_target.py --start 0 --end 76 lora_LSTM whisper- Generate adversarial examples for input source singing voices.
python protect_source.py --start 0 --end 3000 lora_LSTM whisper- Compute the target speaker similarity of undefended output singing voices.
python cal_speaker_sim_undefended.py --start 0 --end 76 --attack_flag all_singers-10_voices --attack_flag_2 all_singers-10_voices resnet18_veri- Process the ground-truth lyric for computing lyric word error rate.
cd wenet/examples/wenetspeech/s0
sh test_WER_CER_args.sh None in OpenSinger- Compute the lyric word error rate of undefended output singing voices.
cd wenet/examples/wenetspeech/s0
sh test_WER_CER_args.sh all_singers-10_voices out inference-all_singers-10_voices- Obtain the defended output singing voices (dual prevention).
python few_shot_svc.py -start 0 -end 76 -type adver-backtrack-lr=0_0002 -source_type backtrack-lr=0_0002 -dataset opensinger -w_start 0 -w_end 10000- Compute the target speaker similarity of defended output singing voices.
python cal_speaker_sim_defended.py --start 0 --end 76 --attack_flag adver-backtrack-lr=0_0002 --attack_flag_2 backtrack-lr=0_0002-adver-backtrack-lr=0_0002 resnet18_veri- Compute the lyric word error rate of defended output singing voices.
cd wenet/examples/wenetspeech/s0
sh test_WER_CER_args.sh backtrack-lr=0_0002-adver-backtrack-lr=0_0002 out inference-backtrack-lr=0_0002-adver-backtrack-lr=0_0002The step 5, 6, 7, 11 may take quite long. You can reduce the number of (target_singer, source_song) pairs in select-target_speakers-source_speeches-des.yaml.
-
High/Low Hierarchy Multi-Target Loss
-
Frame-level interaction reduction-based (FL-IR) loss
-
Encoder ensemble
-
The English dataset NUS-48E
-
More SVC models
-
Metrics: success reduction rate, SNR, PESQ
-
Robustness evaluation
If you find our work or code useful, please cite our paper as follows:
@inproceedings{SongBsAb,
author = {Guangke Chen and
Yedi Zhang and
Fu Song and
Ting Wang and
Xiaoning Du and
Yang Liu},
title = {SongBsAb: {A} Dual Prevention Approach against Singing Voice Conversion
based Illegal Song Covers},
booktitle = {32nd Annual Network and Distributed System Security Symposium, {NDSS}
2025, San Diego, California, USA, February 24-28, 2025},
year = {2025},
}