0
$\begingroup$

Previous researchers have used techniques like Denoising using Spectral Subtraction method and calculating Short Time Fourier Transform (STFT) by dividing the audio data into fixed size chunks and then calculating the frame spectrogram for each of these chunks.

The image below shows how the author has pre-processed his data by manually extracting the frame and calculating it's spectrogram after applying the above-mentioned methods.

What pre-processing techniques exist for such kind of audio data where you need to use the spectrogram images for developing a CNN model, given that all audio files will be of varying length and bit-rates?

pre-process

$\endgroup$
0
$\begingroup$

The best way is to split the audio files (or live input) into analysis windows of fixed size. Then combine each window prediction into a file-level prediction. The classifier can be trained per analysis window. Or per audio file using Multiple Instance Learning.

The window size must be set based on the event of interest, normally a bit longer than the typical event duration. If some files are shorter then the analysis window, pad the input. Based on your image, 1 second may be a reasonable starting point.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.