Voice modeling for music information retrieval and human-computer interaction

点击翻译按钮获取中文摘要

The human voice is an expressive and eloquent musical instrument.  Song and other human-instrument music is ubiquitous in many cultures and is one of the primary forms of modern entertainment.  With the rise of internet music sharing and sales, interacting with music is becoming a large research topic, but current techniques often concentrate on timbre, rhythm, and other characteristics that music listeners would consider the "background" of the music. Much of the "content" of current popular music is contained in the singer's lyrics, expression and style. When pitch is addressed, it is usually instrumental pitch, and when lyrics are considered, they are usually available as text.One of the primary goals of the music information retrieval community is to develop systems whereby a user can hum a tune and the system will identify that tune from a database of possible music.  This difficult task is compounded when users want to add more information by singing lyrics.  One of the aims of this work will be to develop feature models, signal processing techniques, and classification systems that allow for the transcription of human singing.People don't only listen to music, they create it as well. Many people claim that they cannot carry a tune, and yet are capable of humming or whistling along with the radio. This research has the potential to provide amateurs, untrained singers and even non-singers with the tools to interact with music in new and less threatening ways, and gain confidence with music. Professional singers will be provided with tools that make use of their training to facilitate writing of music, to hone their skill, and to interact with music in challenging and intriguing ways.  In the long term, deep and accurate models of human song may lead to a computer that sings to us.