100
that life’s a bore so full of the superficial
that life support of all of the superficial
A thousand lights had made me colder
A thousand lies have made me colder
Hear him with the women just around midnight
Hear him whip the women just around midnight
Burned out dealer to the teachers pet
Burnouts deal it to the teacher’s pet
Cause waiting at the answer to his questions is a definite blow
Persuade him that the answer to his questions is a definite no!
Figure 5.2. A list of pairs of transcriptions. Each pair gives two different tran-
scriptions of the same part of a song that we encountered in the retrieved lyrics.
in the extracted text fragments. Varieties occur as a result of mishearings, typo’s
and the use of abbreviations such as
repeat chorus.
In Figure 5.2 we give a number of examples of transcriptions we encountered.
We next want to align the extracted text fragments to easily visualize the dif-
ferences and to come up with a most probable version of the lyrics. This version is
constructed using the lyrics identified on the web. The final version thus does not
need to occur as such on the web.
Aligning multiple sequences is known to be an NP-hard problem [L. Wang &
Jiang, 2004], for many sensible choices of the objective function such as the sum-
of-pairs objective function. For a given alignment of
n sequences the sum-of-pairs
objective function simply sums up the score of all sequence pairs in the alignment.
Several approximation algorithms have been proposed in the literature, e.g. [L.
Wang & Gusfield, 1997; L. Wang, Jiang, & Lawler, 1996].
We choose the following approach. We first select a reference sequence and
optimally align each of the other sequences with this reference sequence. Next we
combine all these individual alignments into a single alignment of all sequences.
As reference sequence we simply choose a sequence of maximum length, as we
expect this sequence to give a complete transcription of the intended song. Shorter
sequences may not include repeating parts explicitly or miss the beginning or end
of the song.
Aligning a Pair of Lyrics
We align a pair of lyrics on the word level. To realize this, we opt for a dynamic
programming approach where we align a pair of strings
S
1
and
S
2
in a 2
× l matrix
Dostları ilə paylaş: