MFCC based hybrid fingerprinting method for audio classification through LSTM

In this paper, a novel audio finger methodology for audio classification is proposed. The fingerprint of the audio signal is a unique digest to identify the signal. The proposed model uses the audio fingerprinting methodology to create a unique fingerprint of the audio files. The fingerprints are created by extracting an MFCC spectrum and then taking a mean of the spectra and converting the spectrum into a binary image. These images are then fed to the LSTM network to classify the environmental sounds stored in UrbanSound8K dataset and it produces an accuracy of 98.8% of accuracy across all 10 folds of the dataset.