摘要
变异情况对语音的影响是导致语音识别系统性能下降的原因之一。一般情况下变异语音数据采集困难,获得的训练数据量少,这样即使测试环境和训练环境都相同,识别性能也不理想。利用自适应算法可以解决这类问题,它采用少量的测试环境数据进行训练,以达到训练模型和测试数据匹配的目的,保证系统良好的识别性能。MAP算法是常用的自适应算法,大多应用于话者自适应环境,该文尝试将其应用于变异语音识别系统中,并通过对该模型做相应改进获得了较好的识别结果。在小词表特定人应力变异的识别实验中,分别用非特定人模型和改进的特定人模型作为初始模型,应用MAP算法,系统识别率均有明显提高,与基本识别系统相比,在10遍自适应数据前提下,识别率分别提高了15.84%和15.97%,最好的识别率达到85.56%和90.42%。
The perform an ce of a speech recognition system often degrades under stress condition.For the difficulty of stressful speech collection,although tested and trained in the same conditions,speech recognizer performs imperfect with sparse data.With s mall amount of data adapting to testing environment ,adaptation algorithms ar e good ways to ensure good system performance.And among these adaptation method s MAP algorithm is a regular choice.In this paper MAP algorithm is explored in the stressful speech recognition system.Based on the improved speaker dependen t model it shows better results.Experiments are conducted in speaker dependen t isolate system under G-force.Selecting speaker independent and improved sp eaker dependent models as prior models respectively,the recognition rates usin g MAP algorithm are both improved impressively.Compared with baseline system,t he increases of10times adaptation data are15.84%and15.97%,and the best results are85.56%and90.42%respectively.
出处
《计算机工程与应用》
CSCD
北大核心
2004年第5期42-44,共3页
Computer Engineering and Applications
基金
国家自然科学基金资助项目(编号:60085001)
关键词
语音识别
变异语音
MAP算法
speech reco gnition,stressful speech,MAP algorithm
