Marko Helen

This paper proposes measures for estimating the similarity of two audio signals, the objective being in query-by-example. Both signals are first represented using a set of features calculated in short intervals, and then probabilistic models are estimated for the feature distributions. Gaussian mixture models and hidden Markov models are tested in this study. The similarity of the signals is measured by the congruence between the feature distributions or by a cross-likelihood ratio test. We calculate the Kullback-Leibler divergence between the distributions by sampling the distributions at the points of the observations vectors. The cross-likelihood ratio test is evaluated using the likelihood of the first signal being generated by the model of the second signal, and vice versa. Simulations were conducted to test the accuracy of the proposed methods on query-by-example of audio. On a database consisting of of speech, music, and environmental sounds the proposed methods enable better retrieval accuracy than the existing methods.

Various methods for query by example for audio signals are discussed in this paper. The query by example aims at automatic retrieval of audio excerpts similar to a user-provided audio sample from his/her personal audio database. Methods based on hidden Markov models, feature histograms, and likelihood ratio test are discussed. A class based approach was adopted in defining the concept of similarity between two audio samples. For this reason two databases were constructed for the simulations. Experiments were carried out by using a high quality audio database and an audio database collected with a mobile phone. Considering the difficulty of the task the results are quite encouraging also from the future research point of view