基于MMSE谱减算法的农产品市场信息语音识别技术  

Speech Recognition of Agricultural Market Information Based on MMSE Spectral Subtraction

作　　者：许金普[1,2] 

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

出　　处：《河南农业科学》2015年第5期156-160,共5页Journal of Henan Agricultural Sciences

基　　金：国家自然科学基金项目(61271364)

摘　　要：为解决传统的便携式农产品市场信息采集设备操作不便,易受使用环境影响等问题,提出利用语音识别技术采集信息,以增加操作界面的灵活性。为增强语音识别的抗噪声鲁棒性,针对农产品市场信息采集的特殊工作环境,采集到20男20女语音训练集材料。首先利用最小均方误差(MMSE)谱减法进行前端带噪语音增强,得到增强后的语音信号,然后提取其MFCC特征用于HMM声学模型的训练;声学识别单元采用上下文相关的三音子模型,模型训练过程中采用了决策树状态聚类和增加高斯混合分量的策略,以提高模型的精确度。在3处不同环境不同信噪比情况下对训练出的模型进行测试,结果表明,MMSE谱减算法处理后的语音识别率比基本谱减法(SS)、多带谱减法(MB)有明显的提高,特别是在较低信噪比情况下更为明显。In order to solve the inconvenient operation of traditional agricultural market information collection portable device,and other issues such as susceptiblity to environment,we proposed the use of speech recognition technology to collect information in order to increase the flexibility of the operator interface. To enhance the robustness of speech recognition under the special working conditions of agricultural market information collection, we collected 20 males and 20 females voice training set material. Firstly,a noise spectral subtraction of the minimum mean square error （ MMSE） at front-end was adopted,and then the MFCC features were extracted from enhanced speech signal for training HMM acoustic model;Acoustic identification unit was a context-sensitive Triphone model. In the training process,decision tree state clustering and increasing Gaussian mixture component strategy were adopted to improve the accuracy of the model. The trained models were tested in three different environments by different SNR speech sentences, and the results showed that the recognition rate of this method was more significantly improved than that of the basic spectral subtraction （ SS ） , multi-band spectral subtraction （ MB） ,especially at low SNR.

关 键 词：农产品市场信息 语音识别 谱减算法 最小均方误差 

分 类 号：S126[农业科学—农业基础科学]