Study of reverberation robust pitch estimators for the singing voice

Making machines understand us has been a challenging issue in the later years. Although reverberation is an omnipresent phenomenon in our daily lives, computers are still not prepared to handle it correctly. A study to help machines overcome reverberation when estimating fundamental frequency is presented. The study focuses on the singing voice since it is the form of human expression with more complex fundamental frequency contours. There have been selected four fundamental frequency estimation algorithms (YIN, TWM, SAC, MELODIA) common for this task in dry conditions. The study evaluates them following the MIREX Audio Melody Extraction evaluation criteria. First, the algorithms are evaluated in dry conditions and different reverberant conditions. It is shown how an increasing reverberation time supposes an increasing loss in accuracy for all algorithms. Besides, MELODIA exhibits a special robustness compared to its competitors. Then, we try to improve fundamental frequency estimators’ performance using different de-reverberation methods (NML, NMF, ITD) as preprocessors. Only NML succeeds in such a task for all algorithms except MELODIA, which keeps performing the best. Anyway, it demonstrates that de-reverberation methods can be used to improve fundamental frequency estimators’ results in reverberant conditions. Finally, the insights of the study results are analyzed. In order to exemplify how the results of this study can be used to improve algorithms’ accuracy, a proof-of-concept algorithm (MIX) is presented. MIX combines MELODIA with SAC and NML dereverberation. It has a general improvement in accuracy of 2% in reverberant conditions and, in addition, it performs as good as the best algorithms in dry conditions: 91% overall accuracy