Improved time-frequency filtering using an STFT analysis-modification-synthesis method

Time-frequency representations provide a convenient method for analyzing signals whose frequency content changes with time, and specifically those signals with components that overlap in time and/or frequency. In this paper we suggest a technique for optimally concentrating components of a signal in the time-frequency plane via a particular choice of analysis window for the short-time Fourier transform (STFT), as well as an efficient method of optimal least-squares synthesis of a component of interest from the STFT. Using the described synthesis method in conjunction with an optimally localized STFT yields a very accurate estimate of a component that may not otherwise be separable from other components or noise in the signal