Hidden Markov Model-based Audio Retrieval

As one component in multimedia, audio signals are filled in the world, which greatly enrich our semantic apperception and acquisition in information society. However, the current way to get information is mainly based on the vision, especially the text. The retrieval based on the audio information is ignored. So, an audio retrieval system is presented in this paper. Depending on time-domain and frequency-domain features: short-time energy, short-time zero-crossing rate, short-time energy-frequency value and mel-coeffients, audio streams are segmented into six classes: commercial, anchorperson, weather forecast, football match, music and drama. Then, an audio retrieval system based on HMM and BP neural network is presented since HMM can simulate stochastic time series data quite well and ANN has many advantages such as parallel processing ability, powerful discriminating ability etc. Based on gradient descent, traditional BP algorithm has a slow operating speed, so an improved BP algorithm is presented in this paper to improve the recognition speed. Experimental results showed its validity.