基于数据模拟和HMM自适应的电话语音识别研究  

A Study on Telephone Speech Recognition Based on Data Simulation and HMM Adaptation

作　　者：左国玉[1] 刘文举[1] 阮晓钢[2] 

机构地区：[1]中国科学院自动化所模式识别国家重点实验室,北京100080 [2]北京工业大学电子信息与控制工程学院,北京100022

出　　处：《计算机工程与应用》2003年第27期29-31,129,共4页Computer Engineering and Applications

基　　金：国家自然科学基金(编号:60172055);中国科学院自动化所领域前沿项目(编号:1M02J05)

摘　　要：该文研究了基于数据模拟方法和HMM(隐马尔科夫模型)自适应的电话信道条件下语音识别问题。模拟数据模仿了纯净语音在不同电话信道条件下的语音行为。各基线系统的HMM模型分别由纯净语音和模拟语音训练而成。语音识别实验评估了各基线系统HMM模型在采用MLLR算法(最大似然线性回归)做无监督式自适应前后的识别性能。实验证明,由纯净语音转换生成的模拟语音有效地减小了训练语音和测试语音声学性质的不匹配,很大程度上提高了电话语音识别率。基线模型的自适应结果显示模拟数据的自适应性能比纯净语音自适应的性能最大提高达到9.8%,表明了电话语音识别性能的进一步改善和系统稳健性的提高。This paper addresses the problem of speech recognition under telephone channel conditions using data simulation method and HMM(Hidden Markov Model)adaptation.The artificially recreated telephone data from clean speech simulates various speech behaviors in telephone channel situations.Recognition experiments are performed to evaluate the performances of the HMM models derived either from clean speech or from simulated ones before and after unsupervised MLLR(Maximum Likelihood Linear Regression)adaptation.The experiments show that the simulated data can reduce the acoustical mismatch between training conditions and testing conditions and improve telephone speech recognition rates in a great degree.The adaptation results of baseline models report that the simulation models give an additional9.8%word recognition rate higher than those of the clean model.New improvements are shown in the recognition performances after HMM adaptation and in the robustness of recognition system.

关 键 词：数据模拟 电话语音识别 HMM自适应 MLLR 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]