This thesis is concerned with the development of techniques that facilitate the effective implementation of capable automatic chord transcription from music audio signals. Since chord transcriptions can capture many important aspects of music, they are useful for a wide variety of music applications and also useful for people who learn and perform music.;A comprehensive and systematic analysis of state-of-the-art approaches developed for automatic chord recognition systems is provided. Through this analysis, this thesis attempts to discover the most important and influential factors affecting the performance of automatic chord recognition systems.;The findings from this analysis serve as a foundation for developing a number of other techniques presented in this thesis. First, a novel feature smoothing technique based on repeated patterns in music is proposed for overcoming the limits of conventional smoothing techniques. Second, a new discriminative training method for HMM-based chord recognition systems is introduced, and its potential as an alternative to the mainstream generative approach in chord recognition is examined. Third, a new feature extraction approach and a new modeling approach are developed for handling a large number of chord types. The resulting system shows advantages in speed and accuracy compared to existing large-vocabulary systems.;The final outcome of this thesis is a web-based chord transcription system that allows users to convert their own audio files to human readable chord transcriptions. This system is open to the public, and can directly assist anybody who wants to transcribe the chords of a song but has difficulties in recognizing them by ear alone.
展开▼
