Key Technology And Implementation Of Audio - Based Retrieval System Based On Content

With the rapid development of the Internet and multimedia, people enter the era of information explosion. How to retrieve the useful information quickly and accurately in the mass of information has become an urgent need. Audio retrieval as an important branch of information retrieval technology has achieved rapid development and become a research hotspot. Meanwhile, with the rapid development of intelligent terminals, audio retrieval needs new requirements. Due to the complexity of intelligent terminals environment, the query often has more noise, which means the system needs high robustness. In order to facilitate communication between the terminal and the server, the data of feature should be much smaller.This paper introduces content-based audio retrieval technology, and the main work and research results include the following aspects:(1) Analysis and study a variety of audio features, such as Mel frequency cepstral coefficients(MFCC), Chroma melody feature, a variety of audio fingerprint features and so on. The results shows audio fingerprint has higher robustness. We improve the Shazam’s audio fingerprint, so that it has good noise immunity, and the data of features is small.(2) Analysis and study some of the common audio classification algorithms, including Dynamic Time Warping, Gaussian Mixture Models, Hidden Markov Models, Support Vector Machines. We use these classifiers to classify the audio features in order to retrieval. Because the time complexity of these classifiers is large, the training of classifiers costs lots of time with the mass of audio data, which makes the performance of classifiers bad. In the retrieval system, we use hash algorithm to classify features.(3) After the classification, the audio data has been fully structured. We use inverted index to retrieve the query. Unlike hash index, it supports massive data, and we can directly use existing systems, such as Lucene.(4) Study audio matching algorithm, and use the improved edit distance as precise matching. It has a faster processing speed after several optimizations.