Statistical humming recognition and theme finder for query by humming systems

Many people have experience of trying to find a piece of music in music stores, and they only have salient tunes in their mind. Sales people usually have no idea about what the tunes are. However, humming and singing provide the most natural means for content-based retrieval from the music database. In this research, a query by humming system is proposed, and two important components in such a system are examined in this dissertation in detail.; The first component is a statistical humming recognizer that transcribes what are hummed by people. Statistic speech recognition techniques are used. The segment of a note in the humming waveform is modeled by a hidden Markov model (HMM) while data features such as pitch measures are modeled by a Gaussian mixture model (GMM). Preliminary recognition experiments achieve an overall recognition rate of around 80%.; The second component is a theme finder. Themes give highlights of a music piece, and can be used to index music database. Short melodies that repeat many times in a music piece are usually referred as the themes of the music piece. Finding themes is the same as finding repeating melodies in music pieces. A theme indexed music database facilitates the query process. Two dictionary based compression techniques, Lempel Ziv 78 and sliding window techniques, are used to find repeating melodies from music pieces. Experiments performed on a popular music database of MIDI files demonstrate that the proposed algorithms extract repeating melodies effectively with a speed of four times faster than that of the traditional linear search approach.; Lastly, advanced humming processing techniques are discussed. Four topics are covered. They include: an enhanced humming recognition technique, music language modeling and its application, a new humming database, and development tools. An expansion of the humming recognition system's structure and the definition of new HMMs note models are experimented. The potential of applying an N-gram based music language model, which is built upon extracted repeating patterns, to the proposed humming recognition system is demonstrated. A humming data collection protocol and development tools are also provided.