We propose an adaptive noise estimation and reduction algorithm which is capable of reducing additive noise from the noisy speech signals with low SNR values. The algorithm uses Modulated Complex Lapped Transform (MCLT) to estimate the power spectrum of input signal. The noise is estimated continuously from the spectrum using time-frequency dependent smoothing factor and tracking spectral minima. The gain function is then estimated using the smoothed a priori SNR value for the current frame instead of the previous frame using two-stage wiener filters. This method is simple to implement and greatly suppresses the residual musical noise as well as delay, providing consistent speech quality improvement across all SNRs and on average, nearly 0.13 Perceptual Evaluation of Speech Quality (PESQ) improvements.