Playback detection using machine learning with spectrogram features approach

This paper presents 2D image processing approach to playback detection in automatic speaker verification (ASV) systems using spectrograms as speech signal representation. Three feature extraction and classification methods: histograms of oriented gradients (HOG) with support vector machines (SVM), HAAR wavelets with AdaBoost classifier and deep convolutional neural networks (CNN) were compared on different data partitions in respect of speakers or playback devices: for instance with different speakers in training and test subsets. The playback detection systems were trained and tested on two speech datasets S1 and S2 manufactured independently by two different institutions. The test error for both datasets oscillates about the level of 1% for HOG+SVM and even below it for CNN in bigger S1 base. In cross validation scenario in which one base was used for training and second base for the test the results were very poor what suggests that the information relevant for playback detection appeared in each base in different way.