An adaptive time-frequency analysis scheme for improved real-time speech enhancement

An adaptive time-frequency analysis scheme is proposed for improved real-time speech enhancement. The proposed scheme uses a filtering of the short-time Fourier transform (STFT) to obtain an adaptive resolution and is computationally efficient, since it uses fast Fourier transforms (FFTs) in the analysis and synthesis filters. Unlike previously suggested methods, the proposed method allows the time-frequency resolution to be chosen independently for each frequency bin. Perfect reconstruction is achieved by calculating the speech enhancement gains based on the adaptive analysis and then applying these to a STFT which is then filtered to allow the causal part of the gains to pass through the synthesis filter. The proposed method is shown to have superior performance compared to a fixed resolution STFT scheme with an equal time delay of 10 ms.