语音识别错误对翻译性能的影响分析

Analysis of the effects of speech recognition errors on translation performance

传统的机器翻译模型均基于无噪声环境,即输入的数据是无错误的.但在实际同声传译中,语音识别不可避免会存在错误,这些错误在机器翻译过程中会直接影响其他内容的翻译.因此,统计分析语音识别错误的种类及产生的影响对提高机器翻译的鲁棒性具有指导意义.为了模拟真实语音识别错误,本文通过人工朗读NIST汉英实验测试集,并采用讯飞语音识别系统获取其语音识别结果进行统计分析,主要包括:1)语音识别错误的词性分析;2)语音识别错误的类型分析;3)语音识别错误对翻译性能的影响;4)语音识别错误对其他词翻译的影响.得出的主要结论为:名词和动词出现语音识别错误的次数较多,人名最易出现语音识别错误;同音异形字的语音识别错误出现次数最多;长度较小的句子在翻译时受到语音识别错误影响的程度更加明显;与语音识别错误词距离更近的词的翻译更易受到影响.

Traditional machine translation models are based on a noise-free environment.In other words,they are based on noise-free input data.However,in practical applications,user input will inevitably endure noises,and these noises will directly affect the translation of other content during the machine translation process.Statistical analyses on this type of noises and their impact can instructively improve the robustness of machine translation.To simulate real speech noises,we have collected results with speech noises from the NIST Chinese-English experimental test set through manual reading and have adopted the iFLYTEK speech recognition system.Furthermore,we have conducted statistical analyses on the noise text,including 1)part of speech analysis of noise,2)analyses of the types of noises,3)the influence of noises on translation performances,and 4)the influence of noise on the translation of other words.Main conclusions can be drawn below.Nouns and verbs are most prone to speech noises;homophone noises appear most frequent;sentences with smaller lengths are severely affected by noises during translation;and words nearer to noise words are more likely affected by noises during translation.