Content Retrieval against large-scale spoken documents based on the integration of speech and language processing

点击翻译按钮获取中文摘要

We conducted the research and the development of spoken content retrieval targeting large-scale spoken documents. Firstly, for the spoken term detection (STD) task, which aimed to detect the position in a spoken document that a given term appeared at, we developed the method that did not require any detection threshold but, instead, outputted the candidates in increasing order of their plausibility. Finally, we achieved about 70 times faster detection at the almost same detection performance than the baseline continuous DP matching. Next, for the spoken content retrieval (SCR) task, which aimed to find the segment in a spoken document that was relevant to a given query topic represented in natural language, we developed the method robust for recognition errors and out-of-vocabularies (OOVs) that made use of STD as its preprocessing. We found that the proposed method was effective for the query including OOVs and worked complementally with the conventional SCR method, which made use of the large vocabulary continuous speech recognition (LVCSR), and that the combination of them improved the retrieval performance.