Robust pitch period detection using dynamic programming with an ANN cost function

WITH AN ANN COST FUNCTION S. Harbeck A. Kie ling R. Kompe H. Niemann E. N oth Univ. Erlangen-N urnberg, Lehrstuhl f ur Mustererkennung (Inf. 5), Martensstr. 3, 91058 Erlangen, F.R. of Germany e-mail: [email protected] ABSTRACT In this paper, a new pitch synchronous F0-algorithm is described. The task of detecting pitch periods in the speech signal is solved with a search for an optimal path through a space of pitch period hypotheses. The search is e ciently implemented by dynamic programming (DP). The DP cost function is computed with an automatically trained arti cial neural network (ANN) which combines the outputs of heuristic functions measuring the similarity of adjacent period hypotheses. With this algorithm a coarse error rate of 4,75% on a German speech database is achieved. It outperforms the DPF0 algorithm, which itselfs outperforms two \conventional" algorithms.