一种基于Gammatone滤波器组的语音识别噪声鲁棒特征  

A Gammatone Filterbank-Based Speech Feature for Improving Speech Recognition

作　　者：马宇飞 陈骁 王荔 陈龙 张晓灿[2] MAYufei;CHEN Xiao;WANG Li;CHEN Long;ZHANG Xiaocan(Key Laboratory of Cognition and Intelligent Technology,Information Sciency Acadamy of China Electronics Technology Group Corporation,Beijing 100086,China;Intelligent Voice Technology Center,The 3rd Research Institute of China Electronics Technology Group Corporation,Beijing 100015,China)

机构地区：[1]中国电子科技集团公司信息科学研究院认知与智能技术重点实验室,北京100086 [2]中国电子科技集团公司第三研究所智能语音技术中心,北京100015

出　　处：《电声技术》2022年第9期97-100,104,共5页Audio Engineering

摘　　要：提出一种应用于噪声环境下语音识别的基于Gammatone滤波器组的语音特征。相较于传统基于滤波器组的语音特征,该特征值将传统的应用于频域傅里叶变换(Fast Fourier Transform,FFT)的滤波器组的频域分布特点转化为缩放系数,直接应用在频域缩放用于生成倒谱系数的基向量。该信号处理方式最大程度地保留了语音信号频域原有细节。在此基础上,所提出的方法还针对噪声信号的特点对特征生成过程中的分帧长度进行了优化。实验验证了该方法在噪声环境下语音识别的鲁棒性。In this paper, a speech feature was proposed for Automatic Speech Recognition(ASR) in noisy environments. The proposed feature was designed based on the Gammatone filterbank. Compared with the conventional speech features in which filterbanks were implemented on Fourier Fast Transforms(FFT) directly. Gammatone filterbank was transformed into scaling vectors to make Gammatone scaled basis vectors with which the proposed speech feature was generated. The detail information of the spectral signal was maintained. Meanwhile, the frame length of the algorithm was optimized based on the property of the noisy speech signal. The proposed feature performed well in noisy environments for ASR.

关 键 词：语音识别 噪声环境 语音特征 

分 类 号：TN912.31[电子电信—通信与信息系统]