To make an improvement of the traditional Bayesian Nonnegative Matrix Factorization (BNMF), an algo-rithm combining the traditional BNMF based speech enhancement with Gaussian mixture model was proposed. It mainly consisted of a training stage and an enhancement stage. In the training stage, the dictionaries of speech, noise and the combined dictionary were constructed respectively by analyzing the training sets of speech and noise. In the enhancement stage, by using Minimum Mean Square Error (MMSE) estimator, the clean part of the speech was finally reconstructed from the noisy speech. Compared with the traditional speech enhancement methods of NMF and BNMF, extensive experiments indicate that this algorithm has lower speech distortion and better noise suppres-sion performance.%为改进传统贝叶斯非负矩阵分解(BNMF)语音增强算法的性能,提出基于高斯混合模型的贝叶斯非负矩阵分解语音增强算法。该算法分为训练和增强两个阶段,训练阶段,对纯净语音与噪声分别进行训练,得到纯净语音字典、噪声字典与联合字典;增强阶段,采用最小均方误差法(MMSE)从带噪语音中重构原始干净的语音,达到语音增强的目的。实验表明,该算法在提高语音质量和抑制背景噪声等方面,均优于非负矩阵语音分解(NMF)算法与 BNMF 算法。
展开▼