Audio Segment Retrieval Using a Synthesized HMM

In this paper, we propose a general approach to audio segment retrieval using synthesized HMMs. The approach allows a user to query audio data of any length by one or more example audio segments and find similar segments. The basic idea of our approach is to first train a theme HMM using the given examples and a general background HMM using all the audio data, and then combine these individual HMMs to form a synthesized “BackgroundTheme-Background” HMM. This synthesized HMM can then be applied to any audio stream as a parser to detect the most likely theme segment. A major advantage of this approach is that it does not assume any predefined segment boundaries as in previous work, thus can be expected to retrieve theme segments with more accurate boundaries. Preliminary experiments with music detection have shown promising results.