A NOVEL TECHNIQUE FOR EXACT DETECTION OF PITCH LOCATIONS

This paper presents a novel technique——fast recursive ladder estimation algorithm for speech signal processing. The formulas and the normalized version of autoregressive models are derived by using operator deeomposision. It is shown that the modeling parameters have very good convergence behavior and fast tracking capability. It is studied how to use the basic parameters, such as likelihood variable, reflection coefficients and prediction error covariance to detect pitch locations, plosive sound and other underlying transitions. A reasonable masking parameter is also given. An effective local maximum algorithm with adaptive threshold is designed which eliminated extra informations and the errors for the detection from the original parameters.
This new pitch detection technique has been used to the simulation design of a variable bit rate vocoder and gives very promissing results.
This technique requires small storage and simple arithmatic operations. Therefore it is suitable for hardware (e.g. VLSI) implementations.