Audio Data Mining Using Multi-perceptron

Abstract--Data mining is the activity of analyzing a given set of data. It is the process of finding patterns from large relational databases. Data mining includes: extract, transform, and load transaction data onto the data warehouse system, store and manage the data in a multidimensional database system, provides data, analyze the data by application software and visual presentation. Audio data contains information of each audio file such as signal processing component- power spectrum, cepstral values that is representative of particular audio file. The relationship among patterns provides information. It can be converted into knowledge about historical patterns and future trends. This work involves in implementing an artificial neural network (ANN) approach for audio data mining. Acquired audio is preprocessed to remove noise followed by feature extraction using cepstral method. The ANN is trained with the cepstral values to produce a set of final weights. During testing process (audio mining), these weights are used to mine the audio file. In this work, 50 audio files have been used as an initial attempt to train the ANN. The ANN is able to produce only about 90 % accuracy of mining due to less correlation of audio data