How to remove false positives in audio classification?

Dear all Audio Experts,

I am trying to detect a particular sound of a footstep from the environment of woods based on the crackling sound of dry leaf by a footstep. I intend not to use any pre-processing (sound filters) since the range of frequencies generated by the footstep are from 0Hz to 50K Hz. The application that is trained using machine learning is giving me many false positives. Also, once a sound is generated by crackling of a leaf (which is used for training of the system) can never be generated again in lifetime. That means, the sound used for testing is so unique and the system has never seen such sound during training. Yet using Timbral Audio features and Low-level descriptors the footstep is detected using SVM and Other classifiers. However, the biggest challenge I am facing is how to get rid of the false positives. Will there be any impact of normalizing the sound levels on features and classification accuracy? Is there any other way to remove the false positives? Kindly help

Answer 1

[Nayan]

Hi

I understand that the problem is with the excessive number of "False Positives" in the classification of footsteps from wood environment sound. False positive indicates that the Machine Learning model somehow classifies the non-footsteps as footsteps. The following can be the reasons for the same

1. Imbalanced dataset: The number of non-footstep data points is proportionately larger than footstep data points. "data-imbalance."
2. Non-characteristic Features: The features used for classification may not capture the distinctive characteristics appropriate for the classification problem at hand.
3. Model Overfitting: Sometimes, the model may be complex enough to fit to noise(outliers) and can classify incorrectly.

Above are a few reasons that can lead to excessive False Positives. I would suggest you look at the mis classified data closely, extract the right features using the "AI for Audio" from the "Audio Toolbox (mathworks.com)" that is distinctive of the classification problem and try to observe if there is a pattern to the miss-classification.

I suggest you refer to the documentation and examples in "Audio Toolbox (mathworks.com)" for a better resolution to your problem.

I hope this helps!

Cheers,

Nayan

[Saurabh Deshmukh]

Dear Nayan,

No words can complete my thankful feeling for your reply. The approaches you have suggested to remove the false positives are really so methodical and scientific. Thank you so much for your reply. I shall one by one try to resolve these from my System. Lets see if the system works . Really thankful to you from the bottom of my heart. I wish I could discuss the system in more depth with you. Please provide me your contact/phone details and will be grateful to you for your time. Really thank you very much again.

Thanks again