Spoken Language Processing for automatic captioning to multilingual contents

研究了双语自动语音识别（ASR），它同时且互补地执行每种语言的ASR。双语ASR使用翻译模型（translation model，TM），并且将其分数添加到在普通ASR系统中使用的声学和语言模型分数。因此，1）TM建模，2）TM得分的有效计算。（3）以英日双语语音对为例，详细研究了双语语音中对应语句的自动检测。给出了合适的语音识别模型以及相应语音检测的效果和存在的问题。我们显著减少了TM分数计算的处理时间，而没有显著降低双语ASR的准确性。

Bilingual automatic speech recognition(ASR), which performs ASRs of each language simultaneously and complementarily, was investigated. The bilingual ASR uses a translation model(TM) and its score adding to acoustic and language model scores which are used in common ASR systems. Therefore, 1) TM modeling, 2) efficient calculation of TM scores. and 3) automatic detection of corresponding utterances in bilingual speech were studied in detail using English-Japanese pair. Suitable TMs and effects and problems of detection of corresponding utterance were shown. We reduced processing time significantly for TM score calculation without significant degradation of bilingual ASR accuracy.