Research On Special Speech Retrieval Technology Through Short-Speech Segments

To make a good use of the sharply increasing audio information, audio retrieval has gradually become the research focus, and special speech retrieval, as an important part of audio retrieval, has attracted growing attention for its wide and promising application. Special speech retrieval refers to finding out the same or a similar audio data from the data base when given a query. Nowadays, there are lots of problems waiting to be solved in the field of special speech retrieval, like poor robustness in S/N and hard to set up index. Generally speaking, this field is still in process of experiment and exploration, lacking in practical technology and system.This paper studied on special speech retrieval in short speech data base in consistent of some key project in information and technology industry of National High Technology Research and Development Program 863. In special speech retrieval, the calculated amount of comparing the distance between query speech and speech segments for indexing is too huge to accept, while the speed of calculating statistic feature is fast. However, this is still not a perfect method, for the indexing accuracy by using statistic features is very low. On the basis of analysis and study on current algorithm, this paper put forward a two-step speech retrieval method which firstly makes a quick select in the data base and then matches accurately, for the purpose of speeding up retrieval and raising the accuracy rate. The details are as follows:1. To put forward a quick histogram filtering algorithm based on the slip-overflow test. According to the relationship of slip distance and short speech file length, the quick histogram filtering algorithm based on the slip-overflow test is put forward on the basis of active histogram retrieval algorithm, which can exclude data different from query speech quickly and accurately.2. To propose Weighted Locally Preserving Projections(WLPP) algorithm for feature vector dimension reduction. In order to improve the accuracy of speech retrieval, the distances between query speech feature vector and speech segment feature vectors are calculated. The calculated amount of these high-dimension vectors is huge, which badly effects retrieval efficiency. Therefore, WLPP algorithm based on manifold learning theory is applied to reduce the dimension of speech feature vectors. The simulation has proved that, this new method can not only ensure the retrieval accuracy after dimension reduction, but also make the retrieval accuracy not sensitive to the reduced dimension.3. To bring up an idea of Two-directional Two-dimensional Weighted Locality Preserving Projections ( (2D)2WLPP ) algorithm. Owing to the redundant information of in-frame and cross-frame features, this paper poses the idea of (2D)2WLPP to decrease the redundant dimension of in-frame and cross-frame. Experiments show that this new way brings in more effective solution to dimension reduction, increasing in retrieval velocity while decreasing in retrieval accuracy compared with WLPP algorithm.4. To design a two-step speech retrieval system and then build a framework according to the preceding research which gives a concrete description on the structure and function of that system and analysis of major parameter in it.