Research On Content-Based Audio Information Retrieval

In addition to visual media, audio is the most important media of multimedia resources. With the number of multimedia that can be used increasing rapidly, how to find them with an effective way, has become an important research topic of multimedia and information retrieval technology. In the content-based audio information retrieval technology, the most basic work is to extract audio features. Different features lead to different result of accuracy in audio classification and audio information retrieval. In addition, the method of matching will also affect accuracy in retrieval.In the research on audio classification, to distinguish between human voice and sound with background music, make a feature called"Ratio of Low Amplitude Signal"(RLAS). And to distinguish between speech sound and singing sound, make a feature called"Scattered Degree of Waveform"(SDOW). RLAS is the rate of low amplitude samplings in all samplings when the sound is not mute. Experimental tests show that the waveform of human voice have more low amplitude signal than that of the sound with background music.SDOW is the degree of scatter about waveform graphic. The waveform graphic of speech sound is more scattered than that of singing sound. Finding out an appropriate approach to measure degree of scatter can be distinguish between speech sound and singing sound. In the field of audio information retrieval, the major studies are the technology of music retrieval by humming and audio retrieval by example. In the research of music retrieval, based on frequency analysis in feature extraction, with voice characteristics, according to the jumping level of the largest loudness musical note in audio frame, to separate musical notes from the audio of humming or whistle tune. And according to the jumping level of prominent frequency band to separate musical notes from the audio of solo singing. And according to the ratio of different notes and the degree of deviation to measure similarity on matching. In the research on audio retrieval by example, make a improvement on the segmentation of frequency and the selection of coefficient in the extraction of Mel Frequency Cepstrum Coefficient by using wavelet transform. And analysis chronological requirement on the matching by segmentation and provide a calculation method on how to judge whether the result meet the chronological requirement.