I am a graduate student in the computer science and
engineering department, the Ohio state university. Currently I am
carrying out a project which is to extract the vocal line (singing) out of a
song, i.e. to remove the backing music. In order to develop such a system, I
need recordings where the vocal and music are separately recorded. e.g., the
vocal is in one track and the music is in another track (different instruments
are not necessarily separately recorded). For all the recordings I can found,
the vocal and the music are mixed and then stored in different channels.
Karaoke is not useful since in Karaoke, the music is stored in one channel and
the mixture of vocal and music, instead of the vocal alone, is stored in the
other channel. I heard that in modern pop music industry, every instrument as
well as vocal line is separately recorded. They are mixed before the release. Since
it is difficult for me to access this kind of resource, I am wondering if there
are other ways to get the recordings.
To my knowledge, there is no common test database for the
task. I intend to build a small one which might be used by other researchers in
the future. The problem I am encountering is that I am not familiar with music therefore
I can not determine representative samples for the database. I understand that
singing, as speech, is extremely diverse, and such database might be beyond me.
Nevertheless, I wish the list could give me some ideas about the construction
of the database. Thank you for your attention.
Computer Science & Engineering
The Ohio State University
U. S. A.