Abstract: A time-frequency dictionary learning approach is proposed to enhance speech contaminated by additive non- stationary noise. In this approach, a time-frequency dictionary is used for noise process modeling and incorporated into the convolutive nonnegative matrix factorization framework. The update rules for speech and noise time-varying gains and speech time-frequency dictionary are derived by precomputing the noise dictionary. The magnitude spectrogram of speech is estimated using convolution operation between the learned speech dictionary and the time-varying gains. Finally, noise is removed via binary time-frequency masking. Experiments indicate that the scheme proposed in this paper gives better enhancement results in terms of quality measures of speech. The proposed algorithm outperforms the multiband spectra subtraction and the non-negative sparse coding based noise reduction algorithm in nonstationary noise conditions.