计算机与现代化

• 模式识别 • 上一篇    下一篇

基于Gabor滤波的语音识别鲁棒性研究

  

  1. (1.兰州理工大学电气工程与信息工程学院,甘肃兰州730050;2.甘肃省工业过程先进控制重点实验室,甘肃兰州730050;
      3.兰州理工大学电气与控制工程国家级实验教学示范中心,甘肃兰州730050)
  • 收稿日期:2017-10-21 出版日期:2018-06-13 发布日期:2018-06-13
  • 作者简介: 缑新科(1966-),男,甘肃天水人,兰州理工大学电气工程与信息工程学院、甘肃省工业过程先进控制重点实验室、兰州理工大学电气与控制工程国家级实验教学示范中心教授,博士,研究方向:模式识别,信号处理; 徐高鹏(1991-),男,陕西榆林人,硕士研究生,研究方向:语音识别。

Research on Speech Recognition Robustness Based on Gabor Filtering

  1. (1. College of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730050, China;
      2. Key Laboratory of Gansu Advanced Control for Industrial Processes, Lanzhou 730050, China;
      3. National Experimental Teaching Demonstration Center of Electrical and Control Engineering, Lanzhou University   of Technology, Lanzhou 730050, China) 
  • Received:2017-10-21 Online:2018-06-13 Published:2018-06-13

摘要: 为了提高语音识别系统的鲁棒性,提出一种基于GBFB(spectro-temporal Gabor filter bank)的声学特征提取方法,并通过分块PCA算法对高维的GBFB特征进行降维处理,最后在多个相同噪音环境对GBFB特征以及常用的GFCC,MFCC,LPCC等特征进行抗噪性能对比,与GFCC相比GBFB特征的识别率提高了5.35%,与MFCC特征相比提升了7.05%,比LPCC特征识别的基线低9个分贝。实验结果表明,在噪音环境下与传统的GFCC、MFCC以及LPCC等特征相比GBFB特征有更优越的鲁棒性。

关键词: 语音识别, 鲁棒性, Gabor滤波, 特征提取, GBFB特征

Abstract:  In order to improve the robustness of speech recognition system, a method of extracting the acoustic features based on GBFB (spectro-temporal Gabor filter bank) is proposed, and the dimension of the GBFB is reduced by the block PCA algorithm. Finally, the feature of GBFB are compared with the feature of GFCC, MFCC and LPCC in different noise environments. The recognition rate of GBFB features is 5.35% better than GFCC features, the recognition rate of GBFB features is 7.05% better than MFCC features. Moreover, GBFB features are 9 dB lower than the LPCC recognition base. The experimental results show that the GBFB features exhibit better robustness than the traditional features of GFCC, MFCC and LPCC in the noisy environment.

Key words: speech recognition, robustness, Gabor filter, features extraction, GBFB features

中图分类号: