ANNOUNCEMENT: voice/music separation examples available

[Avery Wang <avery@xxxxxxxxxxxxxxxxxx>]

I would like to announce the availability of some sound examples from
my thesis, which is entitled "Instantaneous and Frequency-Warped Signal
Processing Techniques for Auditory Source Separation."  The thesis
is still being prepared and should be available later this summer
sometime.  I will make a further announcement.


The technique I use involves an adaptive notch filter/harmonic
frequency-locked loop to track the instantaneous frequency of each
partial belonging to the voice.  I used voices because the technique
works best on continuous, minimally articulated sounds.  Other
musical instruments, such as violin, trumpet, clarinet, etc.,
could be tracked equally well, provided that they are somewhat

continuous.   Segmenting articulated notes is currently problematic.  For
example, a rapid succession of piano notes would cause my algorithm
to become confused.  Automatic segmentation is a future project for
which I have definite ideas.



These sounds are in 44.1KHz, 16-bit, mono NeXT snd format.

Sorry about the bulk.  (At least I collapsed the stereo tracks down to
mono!)

        baritone.diff.snd
        baritone.isolated.snd
        baritone.orig.snd
This is a 4.1 second segment of a baritone and orchestra from the
"Barber of Seville".  The voice and orchestra have been separated
from each other.  The algorithm does well on vibrato.


        soprano.diff.snd
        soprano.isolated.snd
        soprano.orig.snd
Another snippet from the "Barber of Seville."  This 4.1 second
example contains a soprano, baritone, and orchestra.  The soprano
has been separated.  The quality here isn't so great, but then again,
the mixture really is quite complex.


        days.diff.snd
        days.isolated.snd
        days.orig.snd
This is an 8.4 second segment from 10,000 Maniacs' "These are the
days".  Natalie sings in a mix with melody and harmony parts as well as
instrumentals.  The melody part has been separated from the harmony
and instruments.  Despite the articulations, the algorithm manages
to track the melody through the mixture.


        etern.diff.snd
        etern.isolated.snd
        etern.orig.snd
This is a 22 second snippet of Kathleen Battle and Wynton Marsalis
together in "eternal source of light divine" by Handel.  Battle is
accompanied by soft harpsichord plus some trumpet at the end by
Marsalis.  Again, the voice has been separated from the mixture.


These sound files are available via anonymous FTP to
        ccrma-ftp.stanford.edu:
                /pub/Publications/Theses/AveryWangThesis/sound_demos

Enjoy,

    Avery Wang

    Center for Computer Research in Music and Acoustics (CCRMA)
    660 Lomita Dr.          Fax: +1 (415) 723-8468
    Stanford, CA 94305      Tel: +1 (415) 364-5002
    avery@ccrma.stanford.edu