Telephone handset identification by collaborative representations

Recorded speech signals convey information not only for the speakers' identity and the spoken language, but also for the acquisition devices used for their recording. Therefore, it is reasonable to perform acquisition device identification by analyzing the recorded speech signal. To this end, recording-level spectral, cepstral, and fusion of spectral and cepstral features are employed as suitable representations for device identification. The feature vectors extracted from the training speech recordings are used to form overcomplete dictionaries for the devices. Each test feature vector is represented as a linear combination of all the dictionary columns (i.e., atoms). Since the dimensionality of the feature vectors is much smaller than the number of training speech recordings, there are infinitely many representations of each test feature vector with respect to the dictionary. These representations are referred to as collaborative representations in the sense that all the dictionary atoms collaboratively represent any test feature vector. By imposing the representation to be either sparse (i.e., to admit the minimum norm) or to have the minimum norm, unique collaborative representations are obtained. The classification is performed by assigning each test feature vector the device identity of the dictionary atoms yielding the minimum reconstruction error. This classification method is referred to as the sparse representation-based classifier (SRC) if the sparse collaborative representation is employed and as the least squares collaborative representation-based classifier (LSCRC) in the case of the minimum norm regularized collaborative representation is used for reconstructing the test sample. By employing the LSCRC, state of the art identification accuracy of 97.67% is obtained on a set of 8 telephone handsets, from Lincoln-Labs Handset Database.