Tools for Music Information Retrieval and Playing - Intelligent Music Information Systems: Tools and Methodologies - page 139

Figure 26. The comparison between song waveform (gray) and its corresponding MP3 window-switching
pattern (black line). Horizontal axis is the MP3 granule index. The four window types (long, long-to-
short, short and short-to-long) are indexed with 0, 1, 2 and 3 respectively. The metronome template is
represented by black line with -1 peaks (D'Aguanno et al., 2006).

MDCT processing of a subband signal provides
better frequency resolution, it consequently has
poorer time resolution. The quantization of MDCT
values will cause errors that are spread over the
long time window so it is more likely that this
quantization will produce audible distortions.
Such distortions usually manifest themselves as
pre-echo because the temporal masking of noise
occurring before a given signal is weaker than the
masking of noise afterward (Pan, 1995; Zhaorong
2001). This situation appears frequently in music,
with strong drum lines, when the drummer plays
snare or bass drum and then the Window-Switch-
ing Pattern may be used as simple onset detector
with a high threshold.
Wang and Vilermo (2001) propose the Win-
dow-Switching Pattern (WSP) as information to
refine the output of an MDCT\footnote{Modified

Discrete Cosine Transform} coefficient analysis
in a beat tracking contest in order to perform
better error concealment for music transmission
in a noisy channel. The algorithm extracts the
subband MDCT coefficients and than it calcu-
lates any subband energy. A search window is
defined. The basic principle of onset selection is
setting a proper threshold for the extracted sub-
band energy values. The local maxima within a
search window, which fulfils certain conditions,
are selected to be beat candidates. This process
is performed in each band separately. Than the
WSP identified by the encoder is used to refine
the MDCT analysis. The algorithm extracts the
subband MDCT coefficients and then it calculates
any subband energy. A Statistical model selects
the correct beat from the beat candidate set.