Accquiring Background sounds for Audio Deep learning/Machine learning

I am designing a raspberry pi cough detecting machine

I have a cough dataset (5132 samples) and have converted these samples into Mel spectrogram images for my CNN network.

This network is formed and gives me an accruracy score etc.

I have been trying to find background samples but I have only found 4 samples ( Long files that contains white noise). Now if i compare these samples along with my cough dataset it apears that I am overfiiting the data. Is there a way I can use these 4 long samples to generate many samples from it. Say samples 1000 clips from this long clip.

Like in the case of the below link, The background clips are stripped to many clips (1000 clips from just 1 audio file)

https://uk.mathworks.com/help/deeplearning/ug/deep-learning-speech-recognition.html

Or is there any good audio samples dataset which have background sounds ( Many samples included).

If anyone could please guide me through.. My goal is to make the pi detect whether the recorded sound is a cough or a background sound.

Answer 1

[Milan Bansal]

Hi Joenam Coutinho

It is my understanding that you have only 4 audio samples for background noise and you want to generate more samples from these 4 samples in order to avoid overfiiting.

Data augmentation can be used to create more samples from the existing samples to avoid overfitting of the model. For audio samples, "audioDataAugmenter" can be used to create an augmenter object with specified augmentation properties. The desired properties can be specified by Name-Value pairs. Please refer to the below code snippet that illustrates the augmentation of audio samples.

% Read audio file.
[audioInput,fs] = audioread("filename.wav");
% Create augmenter object.
augmenter = audioDataAugmenter("AugmentationMode","sequential","NumAugmentations",250);
% generate new data.
newData = augment(augmenter,audioInput,fs)

Please refer to the documentation link for more information about "audioDataAugmenter" class.

https://www.mathworks.com/help/audio/ref/audiodataaugmenter.html

Hope it helps!