This paper addresses the problem of noise reduction in non-stationary signals. The paper first describes a human physiology based time-frequency (TF) representation (TFHP) using Mel filterbanks. It is then used to improve a noise reduction algorithm that does not require any a priori information about the signal of interest and the noise. This algorithm is efficiently implemented using an original wavelet shrinkage method. The overall method results in an original TF denoising procedure that yields a denoised TFHP (DTFHP). From this representation one can reconstruct a denoised time-domain signal and therefore define a new improved noise reduction algorithm, whose performance is evaluated and compared with other state-of-the-art methods. The performance assessment uses several criteria: (1) signal-to-noise-ratio (SNR), (2) segmental SNR (SSNR) and (3) mean square error (MSE). The results indicate an improvement of up to 4.72 dB with respect to SNR, 2.79 dB w.r.t SSNR and 4.72 dB w.r.t. MSE for a speech database signals corrupted with four different noises. In addition, other applications such as EEG signal enhancement show promising results.