F0 estimation using empirical mode decomposition and complex cepstrum analysis in reverberant environments

Fundamental frequency F0 is an important cue in speech signal processing, but the performance of F0 estimation by using current state-of-the-art methods drastically degrades due to reverberation effects. Therefore, we propose a F0 estimation method that is robust in reverberant environments by using complex cepstrum analysis and empirical mode decomposition (EMD). Speech dereverberation of our method has two parts: one deals with the amplitude cepstrum, and the other the phase cepstrum. In the first part, EMD is used to decompose the averaged amplitude cepstrum of reverberant speech into two groups of intrinsic mode functions (IMFs). The first group is associated with the amplitude cepstrum of the clean speech signals. The second group, which is used to enhance the reverberant speech signals, is associated with the room impulse response (RIR). In the second part, the all-pass phase cepstrum of the target reverberant speech is modified by a particular value related to the reverberation time. F0 is then estimated from the enhanced speech signals. The results showed that the proposed method could estimate F0 more correctly than other methods such as SWIPE and YIN.