Development of Scoring Algorithm for Karaoke Computer Games

Abstract: In a Karaoke computer game, the users receive a score as a measure of their performance. A music recognition system estimates the underlying music notes the users have performed. Many developed approaches use deterministic signal processing. This thesis builds statistical hidden markov models (HMM) to be used in scoring. The HMM-based note models are based on the musical features pitch, accent, zero crossing rate and power. Gaussian mixture models (GMM) are used to determine the probability distributions of the features. A singing database is constructed by an amateur singer who is well-trained in the area of music. During the training stage, an intra-note model is trained using the HTK toolbox. The intra-note model describes the statistical behavior of model states inside a note. Then, a group of note models, covering a wide range of pitch, is built upon the trained intra-note model. Also a recognition network is constructed. During the recognition stage, the test data are decoded with the aid of thecnote models and the recognition network. An experiment compares the proposed approach with the framed-based and note-based approaches. The results show that our approach provides promising results over deterministic approaches.