Information Technology Reference
In-Depth Information
Table 3. Note durations as a percentage of inter-note onset times (INOT) in Experiment 1
Subject
Percentages, ratio of (Duration / INOT), grouped by INOT value
Group
0 < 150ms
150 < 300
300 < 600
600 < 1200
> 1200
ALL
MUSI
90.35
83.56
85.89
77.13
59.99
78.78
BTWN
92.09
79.40
80.26
60.86
55.04
72.13
NONM
91.61
78.28
79.09
63.79
37.20
70.59
ALL
91.43
80.46
81.75
67.75
53.60
74.08
Among the groups, musicians averaged 79%
overall, while BTWN and NONM averaged 72%
and 71%, respectively. The MUSI group score
was pulled down by one subject who scored only
60%; otherwise, its average would have been over
82%. The largest difference among groups was
seen with notes in the (600-1200ms) range.
When comparing the results of one subject
group against another for a given song, we were
also surprised that the group exhibiting the small-
est standard deviation value changed from one
song to the next with no consistent pattern we could
discern. For instance, song d04 (Beethoven's
Fifth) is nearly always heard in the same key, and
we expected musicians in particular would tend
toward reproducing the notes in the key in which
it is usually played, but the MUSI group showed
more variability in starting pitch choice than the
NONM group. Alternatively, the starting pitches
chosen for the folk song Yankee Doodle (d08) were
most consistent with the MUSI group.
While the experiment showed that musicians
in general have higher variability in their choices
for starting pitch, we see from the inconsistent
results of Table 2 that we cannot make any predic-
tions about how musicians or nonmusicians will
choose their pitches for any given song. Thus,
providing information about starting pitch to
any given song-matching algorithm is unlikely
to improve search accuracy.
The large observed difference between note
durations and INOTs from our subjects is very
significant. While many songs do contain rests,
most of the gaps between the time a note was voiced
and the corresponding INOT were not intended
to be rests, and the corresponding song transcrip-
tions in the database lack rests at most of these
points. This result clearly points to a requirement
that any algorithms involving duration informa-
tion must consider INOT as the more relevant
measure of the user's intended duration values.
It furthermore reduces confidence that searching
Discussion
The data for this experiment show that both mu-
sicians and nonmusicians have some proficiency
in choosing starting pitches so that they have a
better likelihood of remaining in a comfortable
singing range when vocalizing subsequent notes.
The results confirm the general conclusions of
Halpern (1989) and of Levitin (1994) that the
choice of starting pitch for different subjects is
dependent upon the song being sung and is not
merely due to chance.
We had hypothesized that songs which required
most of a subject's vocal range would show the
smallest standard deviation, because the choice
of starting pitch determines whether the melody
will stay within the subject's vocal range or not.
We were surprised to find no strong link between
the two. We suggest that the differences in this
case may be due to the natural differences in
comfortable vocal range among subjects; further
study would be required to prove this. In par-
ticular, a formal exploration of the vocal range
of each subject would have given the ability to
make more definitive statements about how well
each subject chose starting pitches with respect
to the subject's range.
Search WWH ::
Custom Search