FLAD

Fake Lossless Audio Detector

Feature

• Determine the audio source format:
  Lossless, AAC, MP3(ex. SoundCloud), Opus(ex. YouTube)
• Not rely on high frequency information:
  Frequencies below 2.4kHz or above 20kHz are not used
• Resists minor noise interference:
  Noisy samples included in training data

Demo

Audio: AAC downloaded from YouTube
Post-processing: EmiyaEngine
Spectrum:

Lossless Audio Checker (not work):

Ours:

Performance

Test set: 200 samples per category (50% with noise)
Result: 798 correct identifications out of 800 samples with 99.75 % accuracy

Dependence

eval only:

• librosa >= 0.8.0
• resampy
• matplotlib
• numpy
• Pillow
• onnxruntime >= 1.5.2

train & test:

• pytorch >= 1.5.1
• efficientnet_pytorch
• livelossplot (option)

export model:

• onnx-simplifier

Usage

generate dataset:

# set input & output path  
audio_root = './music'
ds_root = './dataset'

then python generate.py

train model:

python train.py

test model (in pytorch):

python test.py

export model to onnx:

python export.py

eval model (in onnxruntime):
run python eval.py fake.flac