DeepSpeech

缺少ctcdecoder库解决方案

cd native_client/ctcdecode
make bindings NUM_PROCESSES=8
pip3 install dist/*.whl

安装librosa报错没有llvm

sudo apt-get install llvm-8
export LLVM_CONFIG='/usr/bin/llvm-config-8'
pip install librosa

0.1.1攻击命令

python attack.py --lr 10 --in sample-000000.wav --target "example" --out adversarial.wav --restore_path models

--lr 10 --in commanderSongoriginal/01_original_crafted_fragment_classical_castale.wav --target "call my wife" --out adversarial1.wav --restore_path deepspeech-0.4.1-checkpoint/model.v0.4.1 --iterations 3000

0.4.1

example 攻击命令, 带l2penalty

--in sample-000000.wav --target "example" --out adv.wav --restore_path DeepSpeech/deepspeech-0.4.1-checkpoint/model.v0.4.1 --l2penalty inf

--input_format "flac" --in "/mnt/data/audio_adversarial_examples_deepspeech/Data/from_librispeech/237-134500-0011.flac" --target "ok google adjust my google security settings" --iterations 3000 --out_path "/tmp" --out_audio "/tmp/output_audio.wav" --log_path "/tmp" --restore_path "DeepSpeech/deepspeech-0.4.1-checkpoint/model.v0.4.1" --l2penalty inf

—input_format “flac” —in “/mnt/data/audio_adversarial_examples_deepspeech/Data/from_librispeech/237-134500-0011.flac” —target “ok google adjust my google security settings” —lr 20 —iterations 3000 —out_path “/tmp” —out_audio “/tmp/output_audio.wav” —restore_path “DeepSpeech/deepspeech-0.4.1-checkpoint/model.v0.4.1” —l2penalty inf