Speech signal reconstruction from modified STFT magnitude spectra using homomorphic analysis/synthesis
Mahmud Gadallah, Alaa Rohaiem, Ayman El Gezawy · The International Conference on Electrical Engineering/The International Conference on Electrical Engineering · 2008
This work presents a low complexity method of reconstructing a speech signal from ahomomorphic Cepstrum Coefficients. The homomorphic deconvolution has been themain core to explore the vocal tract envelope characteristics as well as the excitationsignal Pitch frequency in order to simplify the source speech signal as minimum aspossible. This work usually used for the purpose of speech coding in order to minimizethe communication bandwidth. In this work we have avoided some analytical andreconstruction steps that used to cost extra processing time causing extra delay. Thosesteps are usually used in the most existing methods for this kind of work. Those stepsinclude the samples overlapping lag and Iterative phase retaining algorithm. A set ofharmonic sine waves based on the pitch frequency is used as an excitation signal to thevocal tract filter in order to regenerate the speech signal. The intelligibility of thereconstructed signal obtained is 3.5 in MOS of a subjective test, although the speechsignal was still recognizable.