裴俊華 賈海蓉
關(guān)鍵詞: 語音增強; 小字典; 子空間; K?SVD; OMP; 閾值
中圖分類號: TN912.35?34 ? ? ? ? ? ? ? ? ? ? ? 文獻標(biāo)識碼: A ? ? ? ? ? ? ? ? ? ? ? ? 文章編號: 1004?373X(2019)01?0046?05
Abstract: Since the traditional speech enhancement algorithm of small dictionary has the problem of speech distortion for noise elimination, a speech enhancement algorithm based on adaptive small dictionary in subspace domain is proposed. A over?completed small dictionary is constructed by using the eigenvalues of noisy speech signal in the subspace domain to make the dictionary have perfect control mechanism for signal distortion and residual noise, which is possible to minimize the distortion of the signal while eliminating the noise. The [K] singular value decomposition (K?SVD) algorithm is used for sparse representation and dictionary updating for the noisy speech by means of over?complete small dictionary. The correlation threshold and energy threshold are set in orthogonal matching pursuit (OMP) algorithm to adaptively control the reconstruction and iteration times, and reduce the reconstruction time. The experimental results show that, in comparison with the algorithms given in literatures, the new algorithm under different noise backgrounds has higher SNR and PESQ, and can reduce the speech distortion and improve the speech quality.
Keywords: speech enhancement; small dictionary; subspace domain; K?SVD; OMP; threshold
語音增強[1]的目的就是盡可能地從噪聲中提取出純凈語音信號。近年來,基于信號稀疏表示的語音增強算法受到廣泛關(guān)注。稀疏表示[2]是指用盡可能少的非零系數(shù)來準(zhǔn)確表示原始信號。由于使用冗余字典能很好地表示出在稀疏基上近似稀疏的語音信號,對于非稀疏的噪聲不能進行表示,利用稀疏表示的這個特點能夠有效去除信號中的噪聲。K?SVD[3](K?Singular Value Decomposition)算法是最具代表性的一種稀疏表示算法。近年來,文獻[4]提出一種基于頻域上的小字典訓(xùn)練的語音增強算法,文獻[5]提出一種基于Sparse K?SVD學(xué)習(xí)字典的語音增強方法,文獻[6]提出一種基于自適應(yīng)逼近殘差的稀疏表示語音降噪方法。與這些基于頻域的方法相比,信號子空間[7]可通過選取適當(dāng)?shù)睦窭嗜粘俗覽ν],在抑制噪聲的同時減少信號失真。因此,本文把字典訓(xùn)練方法應(yīng)用于子空間域。而小字典易于進行奇異值分解,更能夠體現(xiàn)出語音的局部特性,所以本文提出一種基于子空間域的自適應(yīng)小字典的語音增強算法。在子空間域中用帶噪語音信號的特征值構(gòu)造過完備的小字典,然后將其作為初始字典,對帶噪語音的特征值用K?SVD算法不斷進行稀疏表示和字典更新。其中在OMP[8] (Orthogonal Matching Pursuit)算法中設(shè)置相關(guān)性閾值與能量閾值[9]來自適應(yīng)控制重構(gòu)階段及迭代次數(shù)。
實驗結(jié)果表明,本文算法與原來的小字典語音增強算法相比,語音增強效果更好,且減少了運行時間,證實了新算法的有效性。
注:本文通訊作者為賈海蓉。
參考文獻
[1] YOU H, MA ZHIXIAN, WEI L I, et al. A speech enhancement method based on multi?task Bayesian compressive sensing [J]. IEICE transactions on information & systems, 2017(3): 557?559.
[2] HSIEH C T, HUANG P Y, CHEN T W, et al. Speech enhancement based on sparse representation under color noisy environment [C]// 2016 IEEE International Symposium on Intelligent Signal Processing and Communication Systems. Nusa Dua: IEEE, 2016: 134?138.
[3] RUBINSTEIN R, PELEG T, ELAD M. Analysis K?SVD: a dictionary?learning algorithm for the analysis sparse model [J]. IEEE transactions on signal processing, 2013, 61(3): 661?677.
[4] 李軼南,張雄偉,曾理,等.基于小字典訓(xùn)練的語音增強算法[J].軍事通信技術(shù),2013,34(1):32?38.
LI Yinan, ZHANG Xiongwei, ZENG Li, et al. Speech enhancement based on small dictionary training [J]. Journal of military communications technology, 2013, 34(1): 32?38.
[5] 黃玲,李琳,王薇,等.基于Sparse K?SVD學(xué)習(xí)字典的語音增強方法[J].廈門大學(xué)學(xué)報(自然版),2014,53(1):36?40.
HUANG Ling, LI Lin, WANG Wei, et al. Speech enhancement based on sparse K?SVD dictionary learning [J]. Journal of Xiamen University (natural science), 2014, 53(1): 36?40.
[6] 周偉力,賀前華,王亞樓,等.基于自適應(yīng)逼近殘差的稀疏表示語音降噪方法[J].電子與信息學(xué)報,2017,39(2):309?315.
ZHOU Weili, HE Qianhua, WANG Yalou, et al. Adapted stopping residue error based sparse representation for speech denoising [J]. Journal of electronics & information technology, 2017, 39(2): 309?315.
[7] DAI X Z, YU B, DAI X H. An improved signal subspace algorithm for speech enhancement [C]// 2014 Conference on e?Business, e?Services and e?Society. Berlin: Springer, 2014: 104?114.
[8] YANG H, HAO D, SUN H, et al. Speech enhancement using orthogonal matching pursuit algorithm [C]// 2014 IEEE International Conference on Orange Technologies. Xian: IEEE, 2014: 101?104.
[9] 周偉棟,楊震,于云.改進的正交匹配追蹤語音增強算法[J].信號處理,2016,32(3):287?295.
ZHOU Weidong, YANG Zhen, YU Yun. Speech enhancement by using modified orthogonal matching pursuit algorithm [J]. Journal of signal processing, 2016, 32(3): 287?295.
[10] 華志勝,付麗華.基于塊分類和字典優(yōu)化的K?SVD圖像去噪研究[J].計算機工程與應(yīng)用,2017,53(16):187?192.
HUA Zhisheng, FU Lihua. K?SVD image denoising based on noisy image blocks classification and dictionary optimization [J]. Computer engineering & applications, 2017, 53(16): 187?192.
[11] JOUNG J, SUN S. SCF: sparse channel?state?information feedback using Karhunen?Loève transform [C]// 2015 GLOBECOM Workshops. Austin: IEEE, 2015: 314?319.
[12] NAKAYAMA K, HIGASHI S, HIRANO A. A noise estimation method based on improved VAD used in noise spectral suppression under highly non?stationary noise environments [C]// 2017 European Signal Processing Conference. Glasgow: IEEE, 2015: 2494?2498.