一种基于倒排索引的音频检索方法

An Inverted Index Based Audio Retrieval Method

传统的基于实例的音频检索算法采用顺序索引,检索时需遍历数据库并导致难以忍受的等待时间。针对传统的顺序的索引方法,该文提出基于倒排索引的音频检索算法。该方法首先利用多种音频特征构成的超向量,通过多层音频分割方法将连续音频流分割为特征数值波动幅度小的短时音频段;然后利用事先训练好的音频字典,将短时音频段序列转换为可以表征音频内容的音频字序列,并建立倒排索引;检索时,将用户提交的查询转换为音频字后利用倒排索引无须遍历数据库即可直接定位候选段落,并根据候选段落与查询的内容相似度大小对候选段落进行排序,将排好序的列表作为检索结果。仿真实验以匹配项排名、同类检索结果比例、定位准确性和检索用时4个方面作为评价指标,实验结果显示,该算法能够在平均1.101 s时间内实现92.58%的检索准确率。

Traditional example based audio retrieval algorithms use forward index, with which, retrieval processing need to traverse the whole database, resulting in intolerable response time. This paper proposes an inverted-index based audio retrieval method. Through constructing super-vector comprising several audio features, audio stream is first segmented into short segments with small feature fluctuation; Based on a pre-trained audio word dictionary, short audio segment sequence is then transformed into audio word sequence, from which inverted index is constructed; During the retrieval phase, the query audio sample is transformed into audio words and retrieval is carried out, candidate segments are ranked according to the similarity with the query. Match term ranking, same type ratio, overlap ratio and retrieval time are used to evaluate the performance of the proposed algorithm. The experiment gives 92.58% retrieval precision within average response time of 1.101 s.