Skip to content

A Implementation of SpecAugment with Tensorflow & Pytorch, introduced by Google Brain

Notifications You must be signed in to change notification settings

superhg2012/SpecAugment

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

54 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SpecAugment

This is a implementation of SpecAugment that speech data augmentation method which directly process the spectrogram with Tensorflow & Pytorch, introduced by Google Brain[1]. This is currently under the Apache 2.0, Please feel free to use for your project. Enjoy!

How to use

First, you need to have python 3 installed along with Tensorflow.

Next, you need to install some audio libraries work properly. To install the requirement packages. Run the following command:

pip3 install SpecAugment

And then, run the specAugment.py program. It modifies the spectrogram by warping it in the time direction, masking blocks of consecutive frequency channels, and masking blocks of utterances in time.

Try your audio file SpecAugment

$ python3
>>> import librosa
>>> from specAugment import spec_augment_tensorflow
# If you are Pytorch, then import spec_augment_pytorch instead of spec_augment_tensorflow
>>> audio, sampling_rate = librosa.load(audio_path)
>>> mel_spectrogram = librosa.feature.melspectrogram(y=audio,
                                                     sr=sampling_rate,
                                                     n_mels=256,
                                                     hop_length=128,
                                                     fmax=8000)
>>> warped_masked_spectrogram = spec_augment_tensorflow.spec_augment(mel_spectrogram=mel_spectrogram)
>>> print(warped_masked_spectrogram)
'
[[1.54055389e-01 7.51822486e-01 7.29588015e-01 ... 1.03616300e-01
  1.04682689e-01 1.05411769e-01]
 [2.21608739e-01 1.38559084e-01 1.01564167e-01 ... 4.19907116e-02
  4.86430404e-02 5.27331798e-02]
 [3.62784019e-01 2.09934399e-01 1.79158230e-01 ... 2.42307431e-01
  3.18662338e-01 3.67405599e-01]
 ...
 [6.36117335e-07 8.06897948e-07 8.55346431e-07 ... 2.84445018e-07
  4.02975952e-07 5.57131738e-07]
 [6.27753429e-07 7.53681318e-07 8.13035033e-07 ... 1.35111146e-07
  2.74058225e-07 4.56901031e-07]
 [0.00000000e+00 7.48416680e-07 5.51771037e-07 ... 1.13901361e-07
  2.56365068e-07 4.43868592e-07]]
'

Learn more examples about how to do specific tasks in SpecAugment at the test code.

python spec_augment_test.py

In test code, we using one of the LibriSpeech dataset.

Example result of base spectrogram Example result of base spectrogram

Reference

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

About

A Implementation of SpecAugment with Tensorflow & Pytorch, introduced by Google Brain

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%