Check result of some Augmentation techniques

[auliofrank (auliofrank@gmail.com)]

Hi,
I can check the audio output of 5 augmentation techniques with bin/play.py file.
I would like to check the result of some other augmentation techniques before decide to apply them for training.
Please give me some advice.
Thank you in advance.

[othiele (Olaf Thiele)]

I am not quite sure I understand what you want to do. Please explain in more detail and give some examples, and answer

What and how to report if you need support

We are happy for every new user of Mozilla STT. As it is under active development, stuff changes quickly and you might have problems running it. We are here to help and might point to this thread if needed Please do the following before posting: Search this forum or Google for your problem Read the documentation for your version here Check Github issues for more technical problems If you still struggle, feel free to open a new thread in this forum (don’t answer to this post).…

[auliofrank (auliofrank@gmail.com)]

When I use bin/play.py tool I can only experiment with 5 different augmentation techniques (overlay, codec, reverb, resample and volume). I don’t know how to do experiment with some other augmentation techniques such as pitch, tempo, warp, etc.
For example: when I execute the following command
python3 bin/play.py --augment pitch[p=1,pitch=2] --clock 0.0 --pipe ./wav/5055_055.wav
I get a warning:
Warning: Some of the augmentations cannot be simulated by this command.

P/S: Before training with data augmentation, I would like to experiment every augmentation technique and then I could choose the right value for each parameter of every augmentation technique.

[othiele (Olaf Thiele)]

@auliofrank

Warning: Some of the augmentations cannot be simulated by this command.

This is all still in development and probably just not available. You’ll have to read the code and experiment yourself.

[auliofrank (auliofrank@gmail.com)]

Thank you for you quick reply.
Do you know what file or function I should start to read?

[bernardohenz]

When I first implemented the augmentations on the spectrogram, I’ve provided a script for checking the effect of them (link). As you will see, the reconstructed audio is noisy (this is because we cannot reconstruct the original signal only from its magnitude/spectrogram).

Many of the spectrogram augmentations were refactored in the DeepSpeech code, but I believe they should follow similar behavior to the ones in the link.

The best way for you to play with the parameters is to copy the colab to your drive, so you can change it and rerun it until you find the desired params.

[bernardohenz]

In addition, the file you should look for checking the augmentations is https://github.com/mozilla/DeepSpeech/blob/master/training/deepspeech_training/util/augmentations.py

[auliofrank (auliofrank@gmail.com)]

Thank you so much for your information. It saved me a lot of time.