Poincaré pitch marks

A novel approach for pitch mark determination based on dynamical systems theory is presented. Pitch marks are used for speech analysis and modification, such as jitter measurement or time scale modification. The algorithm works in a pseudo-state space and calculates the Poincaré section at a chosen point in the state space. Pitch marks are then found at the crossing of the trajectories with the Poincaré plane of the initial point. The procedure is performed frame-wise to account for the changing dynamics of the speech production system. The system is intended for real-time use, so higher-level processing extending over more than one frame is not used. The processing delay is, therefore, limited to one frame. The algorithm is evaluated by calculating an average pitch value for 10 ms frames and using a small database with pitch measurements from a laryngograph signal. The results are compared to a reference correlation-based pitch mark algorithm. The performance of the proposed algorithm is comparable to the reference algorithm, but in contrast correctly follows the pitch marks of diplophonic voices.