Speaker identification using sparsely excited speech signals and compressed sensing

Compressed sensing samples signals at a much lower rate than the Nyquist rate if they are sparse in some basis. Using compressed sensing theory to reconstruct speech signals was recently proposed, assuming speech signals are sparse in the excitation domain if they are modelled using the source/filter model. In this paper, the compressed sensing theory for sparsely excited speech signals is applied to the specific problem of speaker identification, and is found to provide encouraging results using a number of measurements as low as half of the signal samples. In this manner, compressed sensing theory allows the use of less samples to achieve accurate identification, which in turn would be beneficial in several sensor network related applications. Additionally, enforcing sparsity on the excitation signal is shown to provide identification accuracy which is more robust to noise than using the noisy signal samples.