基于谱减法和特征补偿的农产品价格语音识别  

Agricultural price speech recognition based on spectral subtraction and feature compensation

作　　者：许金普[1,2] 诸叶平[1] 

机构地区：[1]中国农业科学院农业信息研究所,北京100081 [2]青岛农业大学动漫与传媒学院,山东青岛266109

出　　处：《计算机工程与设计》2015年第9期2528-2531,2566,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(61271364);农业系统智能控制与虚拟技术团队基金项目(CAAS-ASTIP-2015-AII-03)

摘　　要：为解决农产品价格信息采集任务中传统设备缺少语音接口,且通用的抗噪声算法效果不佳的问题,提出一种农产品价格采集环境下的抗噪声算法。利用谱减算法实现前端的去噪,提高输入信号的信噪比,但同时产生的频谱畸变和残留噪声会造成新的失配,利用倒谱均值方差归一化方法进行补偿,减小或消除这种失配,将基本谱减算法(SS)以及多带谱减算法(MB)分别与倒谱均值方差归一化方法联合。实验结果表明,与单独使用上述各算法相比,联合后的算法能有效提高系统的识别率,特别是在较低信噪比(0dB-10dB)情况下该效果更为明显。In the process of agricultural price information acquisition, traditional equipment is lack of voice interface, and general antbnoise algorithms have poor performance. To solve this problem, a noise robustness algorithm specifically for agricultural price acquisition environment was proposed. Spectral subtraction algorithm was used to remove most of the noise from the original signal, improving the signal-to-noise ratio （SNR） of the input signal, however, the emerging spectrum distortion and the re- sidual noise caused new mismatch. Cepstral mean and variance normalization was then.used to compensate to reduce or eliminate the mismatch. The basic spectral subtraction algorithm （SS） and the multi-band spectral subtraction algorithm （ME） were respectively combined with the cepstral mean variance normalization. Experimental results show that compared with any one of above algorithms, the combined algorithm can effectively improve the recognition rate of the system, especially in low SNR （0 dB-10 dB） environment.

关 键 词：谱减算法 特征补偿 倒谱均值方差归一化 农产品价格 信息采集 

分 类 号：TN912.3[电子电信—通信与信息系统]