Big Data Day LA 2015 - Deep Learning Human Vocalized Animal Sounds by Sabri Sansoy of Deustch Advertising

Investigated a couple audio based, deep learning strategies for identifying human vocalized car sounds. In one case Mel Frequency Cepstral Coefficients, MFCCs, were used as inputs into a supervised, logistic regression neural network. In a separate case, Short Term Fourier Transforms ,STFT, were used to generate PCA whitened spectograms, which were used as inputs into a supervised, convolutional neural network. The MFCC method trained quickly on a relative small dataset of 4 sounds. The STFT method resulted in a much larger input matrix, resulting in much longer times for converging onto a solution