Audio data retrieval and recognition using model selection criterion

Model selection criterion is an unsupervised technique that can be used to compare statistical distributions of the data. In this paper a new application of model selection criterion are presented. The technique is applied for direct audio search in German broadcast news with the high variability in duration and loudness of the search patterns. Experiments for identification of 14 environmental sounds are carried out using model selection criterion as distance metric. For environmental sounds detection the decision is based on mutual similarity of compared events to the set of reference events. For audio events recognition Latent Semantic Indexing (LSI) is also tested. Approximately 500 audio segments from 14 sound types are used in the recognition test. The experiments show that the applications of model selection criterion for direct audio search, unsupervised environmental sounds analysis and sounds recognition using LSI are effective and accurate.