正在加载图片...
.214. 智能系统学报 第11卷 将MVDA与MFCC特征在自动语音识别系统 5 下进行语音识别实验对比,实验结果如图4。可以 结束语 得出,信噪比较高时,MFCC特征与MVDA特征的识 本文的分析主要基于加性噪声和卷积噪声环境 别率基本相同,但随着信噪比降低,MVDA语音特征 下MFCC特征参数的失真,针对这一问题提出了 的效果更加显著。 MVDA语音特征提取法。分析得出实验效果与语音 基本特征、滤波器的类型均相关。在使用MVDA滤 100 o-MFCC参数 波法后,相较于MFCC语音特征,自动语音识别系统 95 -MVDA参数 在不同性噪比环境下的识别率提高了2.7%~ 90 卧 15.0%。MVDA特征提取可以达到很多复杂去噪算 法的效果,却可以减少系统对计算能力的要求,减小 80f 系统的时延。因此,MVDA后处理法可以在更小的 75 计算代价下提高系统的鲁棒性,具有较高的实际应 105 0 5 10 20 用价值。 躁声强度/dB (a)噪青类型为white 参考文献: 95 [1]PALIWAL KK,BASU A.A speech enhancement method 90 based on Kalman fltering[C]//Proceedings of IEEE Inter- national Conference on Acoustics,Speech,and Signal Pro- cessing.Dallas,USA,1997:177-180. 80 。MFCC参数 [2]GIBSON J D.KOO B.GRAY S D.Filtering of Colored 75 -MVDA参数 Noise for Speech Enhancement and Coding J].IEEE 70 Transactions on Signal Processing,1991,39(8):1732- 5 0 5 10 15 20 1742 噪声强度/dB (b)噪声类型为pink [3]ZELINSKI R.A microphone array with adaptive post-filte- ring for noise reduction in reverberant rooms[C]//Proceed- 95 ings of IEEE International Conference on Acoustics, % Speech,and Signal Processing.New York,USA,1998: 2578-2581. 禁 [4]MYLLYMAKI M,VIRTANEN T.Non-stationary noise mod- 80 el compensation in voice activity detection[C]//Proceed- ·MFCC参数 ings of IEEE International Conference on Signal Processing -MVDA参数 Conference.Glasgow,Scotland,2009:2186-2190. 105 0 5 1015 20 [5]RAMFREZ J,SEGURA J C,BENFTEZ C,et al.Efficient 噪声强度/dB voice activity detection algorithms using long-term speech in- (c)噪声类型为volvo formation J].Speech communication,2004,42(3/4): 95 271-287. [6]CHOWDHURY M,SELOUANI S A,OSHAUGHNESSY D. 90 A soft computing approach to improve the robustness of on- 85 line ASR in previously unseen highly non-stationary acoustic environments[C]//Proceedings of the 11th IEEE Interna- 80 -MFCC参数 tional Conference on Information Science,Signal Processing 75 -MVDA参数 and their Applications.Montreal,Canada,2012:522-527. 70 [7]GUPTA H A,RAJU A,ALWAN A.Non-linear dimension 5 10 1520 噪声强度/dB reduction of Gabor features for noise-robust ASR[C]//Pro- (d)噪声类型为DE ceedings of IEEE International Conference on Acoustics, 图10自动语音识别结果对比图 Speech,and Signal Processing.Florence,Italy,2014: Fig.10 Comparison of automatic speech recognition results 1715-1719. 8 HANSEN J H L.VARADARAJAN V.Analysis and com-将 MVDA 与 MFCC 特征在自动语音识别系统 下进行语音识别实验对比,实验结果如图 4 。 可以 得出,信噪比较高时,MFCC 特征与 MVDA 特征的识 别率基本相同,但随着信噪比降低,MVDA 语音特征 的效果更加显著。 图 10 自动语音识别结果对比图 Fig.10 Comparison of automatic speech recognition results 5 结束语 本文的分析主要基于加性噪声和卷积噪声环境 下 MFCC 特征参数的失真,针对这一问题提出了 MVDA 语音特征提取法。 分析得出实验效果与语音 基本特征、滤波器的类型均相关。 在使用 MVDA 滤 波法后,相较于 MFCC 语音特征,自动语音识别系统 在不 同 性 噪 比 环 境 下 的 识 别 率 提 高 了 2. 7% ~ 15.0%。 MVDA 特征提取可以达到很多复杂去噪算 法的效果,却可以减少系统对计算能力的要求,减小 系统的时延。 因此,MVDA 后处理法可以在更小的 计算代价下提高系统的鲁棒性,具有较高的实际应 用价值。 参考文献: [1]PALIWAL K K, BASU A. A speech enhancement method based on Kalman fltering[C] / / Proceedings of IEEE Inter⁃ national Conference on Acoustics, Speech, and Signal Pro⁃ cessing. Dallas, USA, 1997: 177⁃180. [2] GIBSON J D, KOO B, GRAY S D. Filtering of Colored Noise for Speech Enhancement and Coding [ J ]. IEEE Transactions on Signal Processing, 1991, 39 ( 8): 1732⁃ 1742. [3] ZELINSKI R. A microphone array with adaptive post⁃filte⁃ ring for noise reduction in reverberant rooms[C] / / Proceed⁃ ings of IEEE International Conference on Acoustics, Speech, and Signal Processing. New York, USA, 1998: 2578⁃2581. [4]MYLLYMAKI M, VIRTANEN T. Non⁃stationary noise mod⁃ el compensation in voice activity detection [ C] / / Proceed⁃ ings of IEEE International Conference on Signal Processing Conference. Glasgow, Scotland, 2009: 2186⁃2190. [5]RAMFREZ J, SEGURA J C, BENFTEZ C, et al. Efficient voice activity detection algorithms using long⁃term speech in⁃ formation [ J]. Speech communication, 2004, 42 ( 3 / 4): 271⁃287. [6]CHOWDHURY M, SELOUANI S A, O'SHAUGHNESSY D. A soft computing approach to improve the robustness of on⁃ line ASR in previously unseen highly non⁃stationary acoustic environments[ C] / / Proceedings of the 11th IEEE Interna⁃ tional Conference on Information Science, Signal Processing and their Applications. Montreal, Canada, 2012: 522⁃527. [7]GUPTA H A, RAJU A, ALWAN A. Non⁃linear dimension reduction of Gabor features for noise⁃robust ASR[C] / / Pro⁃ ceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing. Florence, Italy, 2014: 1715⁃1719. [8] HANSEN J H L, VARADARAJAN V. Analysis and com⁃ ·214· 智 能 系 统 学 报 第 11 卷
<<向上翻页向下翻页>>
©2008-现在 cucdc.com 高等教育资讯网 版权所有