Multipitch tracking based on linear programming relaxation and sparsity-based pitch candidate estimation

This paper presents a linear programming approach for tracking fundamental frequencies in acoustic signal that contains multiple speech sources and noise interference. A sparsity-based pitch estimation method is used to obtain pitch candidates for each signal frame. With conventional methods like exhaustive searching, the computational complexity of multipitch tracking grows exponentially with the number of pitch tracks. We propose to use a linear programming relaxation approach to solve the multiple pitch track searching problem. This approach has low computational complexity while it was found to attain global optimal solution with high probability. Experimental results show that the proposed algorithm is more efficient and more accurate than the conventional tracking method, extended dynamic programming.