A new robust pitch determination algorithm for telephone speech

Pitch determination algorithm is a critical part in speech coding algorithm. However, in telephone quality speech, the fundamental frequency is often weak, or even missing, which affects the performance of the pitch determination algorithms significantly. Thus in this paper, a robust time-domain pitch determination algorithm is proposed to specifically tackle this problem. It restores the weak or missing fundamental frequency by applying a nonlinear process to the original speech, and calculate a combined autocorrelation function (ACF) based on both the original speech and the nonlinearly processed speech. Furthermore, considering the particular properties of telephone quality speech, a Pitch Candidate Refinement Function (PCRF) that incorporates a newly introduced parameter Long-Time Average Pitch (LTAP) is used to refine the combined ACF. Finally the pitch candidates are selected based on the combined ACF and a dynamic programming is utilized to track down to the true pitch. Experiments results on Keele pitch database show that the proposed algorithm can reduce the gross pitch error rate of telephone speech by about 3% compared to a traditional ACF pitch determination method.