SpecAugment with numpy & scipy

This is SpecAugment code with numpy and scipy.

In my test environment, this code is about 100 time faster than using tensorflow.

Usage

pip install -r requirements.txt
python3 specaugment.py

if you want to test your test audio, run this command

python3 specaugment.py -i <input file>

Example output

start to SpecAugment 100 times
whole processing time : 0.9117 second
average processing time : 9.12 ms

• original spectrogram

• warped spectrum

• warped + masked spectrum

Reference

1. https://arxiv.org/pdf/1904.08779.pdf