Refining fundamental frequency estimates using time warping

Algorithms for estimating the fundamental frequency (F0) of a signal vary in stability and accuracy. We propose a method which iteratively improves the estimates of such algorithms by applying in each step a time warp on the input signal based on the previously estimated fundamental frequency. This time warp is designed to lead to a nearly constant F0. A refine ment is then calculated through inverse time warping of the result of an F0 estimation applied to the warped signal. The proposed refinement algorithm is not limited to specific esti mators or optimized for specific input signal characteristics. The method is evaluated on synthetic audio signals as well as speech recordings and polyphonic music recordings. Results indicate a significant improvement on accuracy when using the proposed refinement in combination with several well-known F0 estimators.