The aim of this is to see I can classify urban sounds accurately. I used the Urban sound 8k dataset for my training and evaluation. I decided to use Tensorflow for the classification due to it's extensive api and documentation.
To create the Spectogram I performed an FFT on each wav file to produce a frequency intensity image (AKA spectogram). The image frequencies included in the image are capped at 2500 Hz and have a resolution of 50 Hz. This gives an image size of 50 pixels by 50 pixels with each x-axis pixel representing a 0.1 sec section in the song and each y-axis pixel representing a 50Hz segment in the frequency domain.
To train the tensorflow model simply run Audio Classification.py. Once trained it should achieve ~87%. I probably didn't evaluate it properly so it might be a little off. You can then use the prediction script to have it classify 20 hand chosen wav files to see how it does.
It is important to be aware that the library I used to read the wav files does not support 24-bit wav files, so before trying to use this it is recommended that all wav files are converted to 16-bit. I used the below command to convert all the files to 16-bit.
"*.wav" -exec sh -c "ffmpeg -y -i {} tmp.wav; mv tmp.wav {}" \;To use the scripts as written without any changes it is important that the audio is placed into a folder called "urban_audio" without any subfolder or additional files.