Source phone identification using sketches of features

Speech recordings carry useful information for the devices used to capture them. Here, acquisition device identification is studied using `sketches of features´ as intrinsic device characteristics. That is, starting from large-size raw feature vectors obtained by either averaging the log-spectrogram of a speech recording along the time axis or stacking the parameters of each component for a Gaussian mixture model modelling the speech recorded by a specific device, features of reduced size are extracted by mapping these raw feature vectors into a low-dimensional space. The mapping preserves the `distance properties´ of the raw feature vectors. It is obtained by taking the inner product of the raw feature vector with a vector of independent identically distributed random variables drawn from a p-stable distribution. State-of-the art classifiers, such as a sparse representation-based classifier or support vector machines, applied to the sketches yield an identification accuracy exceeding 94% on a set of eight landline telephone handsets from Lincoln-Labs Handset Database. Perfect identification is reported for a set of 21 cell-phones of various models from seven different brands.