A Fast Algorithm Of Audio Retrieval

Speech recognition, general audio signal analysis and content-based audio retrieval, are the three primary researching fields in machine hearing technology. Among these fields, speech recognition which has became mature and stable, is a traditional research focus, moreover the rapid expanding of audio information data makes the audio retrieval a new hotspot gradually in information retrieval researching. Audio retrieval was defined as an approach to retrieve interested audio slice from whole audio data. Compared with speech recognition, audio retrieval deals with more generally wave-formed audio signal(include speech and music, etc), and its achievements widely applied to remote education, health and medical, digital library, environmental monitoring, index or tag of news and entertainment programs.In this dissertation, an algorithm of image-registration-based fast audio retrieval, FAR(Fast Audio Retrieval), is proposed. Firstly, audio data is separated into frames with techniques of short-time analysis, and MFCC coefficients are extracted from them. Secondly, MFCC coefficients are mapped into black-white-scale image. Finally, employing image registration method, template matching is carried out between MFCC coefficients of reference template and that of test template, for computing the matching degree between the two templates, and then output the result as similarity.Experiments show that image-registration-based fast audio retrieval algorithm was parallel with DTW based audio retrieval algorithm, the algorithm proposed in this dissertation has better performance demonstrated by recall-ratio, precision ratio and F-quota, and after employed with image registration method the algorithm has better efficiency than DTW method. Consequently, the FAR algorithm proposed in this dissertation is suitable for content-based audio retrieval approaches.