Automatic Music Classification and Retreival: Experiments with Thai Music Collection

We present the experimental results of classification and retrieval of Thai music using TreeQ (a tree-structured classifier) and LVQ (Learning Vector Quantization) algorithms in this paper. We use the HTK Toolkit in preprocessing acoustic signals including feature extraction from the Thai music collection. The training set consists of 250 songs -50 songs from each of the 5 genres. Training is divided into three phases using all or some of these songs. The test set consists of 10 songs selected from 5 genres which are not included in training. We trained and tested the music classifiers using both TreeQ and LVQ algorithms with varying parameters such as, Number of Codebook (NOC) and pruning thresholds to identify the effects of different parameters and features in the Thai music classification and retrieval. We observed that TreeQ-based experiments yield faster response-times than those of LVQ; and therefore, a TreeQ-based system maybe appropriate for online (real-time) music retrieval tasks. On the other hand, LVQ-based experiments consistently yield better accuracy than those of TreeQ; and therefore, a LVQ-based system may be appropriate in the music classification task since music classification can generally be performed off-line. We also outlined a Relevance Feedback based Music Retrieval System in this paper.