In order to treat the problem that the effect of traditional speech enhancement methods is not satisfactory under the condition of non-stationary noise environment and low SNR(signal to noise ratio), an algorithm of speech enhancement based on probabilistic latent component analysis was proposed. By analyzing and introducing probabilistic latent component analysis, phonic spectrogram was explicitly modeled as a mixture of marginal distribution products and noise was described by a dictionary of marginals. The estimation of the most appropriate marginal distributions was performed using expectation-maximization algorithm, which is used selectively to reconstruct the signal, separating it from noise, and the goal of speech enhancement was achieved. Simulation results demonstrated that the proposed algorithm is more effective in terms of reducing background noise, improving SNR and decreasing speech distortion than traditional speech enhancement algorithms.
