Using sets of combs to control pitch estimation errors
Résumé
We analyze the errors of a Pitch Estimation Algorithm using the Pitch Function (response of the periodicity estimator as a function of the frequency parameter Fc). The estimator's maximum response for a single signal of fundamental frequency F0 is expected to occur for Fc=F0. Actually the pitch function exhibits many secondary peaks which occasionally cause the errors. When several signals are mixed the main peaks do not reliably represent the F0s of the component signals. By taking as periodicity estimator the correlation of the spectrum module with a uniform infinite spectral comb of fundamental frequency Fc we show that each peak corresponds to a particular value of the ratio Fc/F0=p/q (p and q positive integers). It follows that some secondary peaks can be cancelled either by augmenting the comb with intermediary negative teeth, or by setting to zero some of its teeth. These modified combs can be viewed as combinations of uniform combs of different Fcs. The present study aims at precisely defining and combining the modified combs so that the main peaks of the new Pitch Function reliably indicate the F0s of the components. Examples are given on mixtures of voiced segments extracted from natural speech.