No abstract

In this paper we compare two sets of audio features in task of audio pattern searching based on elementary sound models. The rst set of features consist of well-known melfrequency cepstral coef cients together with their rst and second order time derivatives. The second set was chosen from bag of features by particle swarm optimization algorithm and consist of following audio features: line spectral frequencies (LSF), spectral ux (SFX) and zero crossing rate (ZCR). Experimental results performed on AudioDat sound database show improvement of above 18.6 % of average F-measure when using the second selected combination of features.

In the paper, we proposed a novel approach for query-by-example audio patterns searching method inspired by classical phoneme-based word spotting and speech recognition systems, where larger and more complex models are created as a concatenation of pre-trained phoneme sub-models. Unlike most other methods for sound events classification which uses pretrained sound classes, our system has no default limitations for demanded sound event and any sound example can be added into the search space. Using methods of cluster analysis and Viterbi alignment, we created a database of what we call "elementary sound"models. These models serves as elementary units of sound event and gives our system flexibility and versatility in adding new patterns into query. In the paper, the methodology of elementary sound models database creation and its characteristics are presented. Also process of creating and embedding models of new patterns into the model network is described.