AdvSpeech

Environment Setup

for the user:

git submodule update --init --recursive
conda env create -f environment.yml
conda activate advspeech
pip3 install --force-reinstall torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

for dev:

conda env update --file dev-environment.yml --prune -> poetry update

conda-lock lock -f dev-environment.yml -> lock the environment

HOW TO USE

if you only want to protect the audio, you could either provide the info directly

python main.py --input audios/en_sample/libri_5694.wav --reference audios/en_sample/ref_ws.wav

or use the config

python main.py --config audios/en_sample/config.yaml

sample audio is hosted on https://yrymax.github.io/AdvSpeech/sample_web/ dataset: https://huggingface.co/datasets/Renyi444/AdvSpeech

if you want to see the benchmark, you need you install the env for all the external repositories.

• antifake: https://github.com/WUSTL-CSPL/AntiFake
• NISQA: https://github.com/gabrielmittag/NISQA
• TTS: https://github.com/coqui-ai/TTS
• OpenVoice: https://github.com/myshell-ai/OpenVoice
• CosyVoice: https://github.com/FunAudioLLM/CosyVoice
• Wespeaker: https://github.com/wenet-e2e/wespeaker

then run run benchmark_pipeline.py

HOW TO CONTRIBUTE TO BENCHMARK

to add a new matrix -> contribute under Metrics

to add a new adv method -> add a new method under adv_runner

to add a new synthesizer -> add a new class under synthesizer that inherits from Synthesizer

to add a new experiment -> add a new method under BenchmarkPipeline