Revisiting spectral envelope recovery from speech sounds generated by periodic excitation

We propose a set of new accurate spectral envelope recovery methods for speech sounds generated by periodic excitation based on a set of interference-free power spectrum representations. The proposed methods outperform our previous spectral recovery models used in legacy-STRAIGHT, TANDEM-STRAIGHT and WORLD VOCODERs. We introduce several design procedures of paired time widows which remove interferences caused by signal periodicity in the time domain or in both time and frequency domains. In addition to this interference-free representation, we introduce post and pre-processing to improve recovery accuracy around spectral peak regions. We conducted a set of evaluation tests using voice production simulator and natural speech samples. Finally, we discuss the application of the proposed method on revising high-quality VOCODERs.