Comparative study of time–frequency representations for singing voice signals

Time-frequency analysis is a useful method to describe signals from both the time and frequency perspectives. It is especially effective for analyzing nonstationary signals, where the frequency components change over time. For this purpose, various transforms have been developed, such as the windowed Fourier transform, the continuous wavelet transform, and the Stockwell transform. Singing voices have unique characteristics, such as a rich harmonic structure, dynamic changes in formants, and fluctuations in fundamental frequency, which are different from those of speech or other sound sources. To capture these features accurately, more precise and detailed analysis is required. In this talk, we focus on this point and compare how the characteristics of the above transforms and their modified methods affect the time–frequency representations. Furthermore, we show examples of spectrograms and scalograms of singing voice signals obtained by these transforms and explain how harmonic components and variations of the signal are represented.