Identifying and Retrieving of Audio Sequences by using Wavelet Descriptors and Neural Network with User's Assistance

Audio content identifying, indexing, and retrieving are of great importance for the next generation of Internet search. The retrieval of digital audio, based on similarities between feature vectors of querying and stored audio material, is considered. We used wavelet descriptors for characterizing short music sequences and Euclidean distance as an objective similarity measure. Since music sequences may be subjectively classified in different ways, the retrieval process is upgraded by user's feedback and neural network decision. After the first pass, from the initial set of sequences ordered related to objectively measured distances from the query, the user select and mark those subjectively declared as more similar to a query. Selected sequences update the weights of radial basis function (RBF) neural network: moving the RBF center towards selected sequences and changing the width of RBF for obtaining fine-tuning. New ranking of audio sequences is then performing from RBF output. The user can repeat the process in iterative way until obtaining satisfactory results. Our simulations show that the process converges very fast: after a few iterations (1 to 4) selected fragments are just those we needed.