libs/libvorbis-1.3.2/doc/vorbisfile/crosslap.html

122 lines
5.2 KiB
HTML
Raw Normal View History

2017-01-10 04:08:49 +00:00
<html>
<head>
<title>Vorbisfile - Sample Crosslapping</title>
<link rel=stylesheet href="style.css" type="text/css">
</head>
<body bgcolor=white text=black link="#5555ff" alink="#5555ff" vlink="#5555ff">
<table border=0 width=100%>
<tr>
<td><p class=tiny>Vorbisfile documentation</p></td>
<td align=right><p class=tiny>vorbisfile version 1.3.2 - 20101101</p></td>
</tr>
</table>
<h1>What is Crosslapping?</h1>
<p>Crosslapping blends two samples together using a window function,
such that any sudden discontinuities between the samples that may
cause clicks or thumps are eliminated or blended away. The technique
is nearly identical to how Vorbis internally splices together frames
of audio data during normal decode. API functions are provided to <a
href="ov_crosslap.html">crosslap transitions between seperate
streams</a>, or to crosslap when <a href="seeking.html">seeking within
a single stream</a>.
<h1>Why Crosslap?</h1>
<h2>The source of boundary clicks</h2>
<p>Vorbis is a lossy compression format such that any compressed
signal is at best a close approximation of the original. The
approximation may be very good (ie, indistingushable to the human
ear), but it is an approximation nonetheless. Even if a sample or set
of samples is contructed carefully such that transitions from one to
another match perfectly in the original, the compression process
introduces minute amplitude and phase errors. It's an unavoidable
result of such high compression rates.
<p>If an application transitions instantly from one sample to another,
any tiny discrepancy introduced in the lossy compression process
becomes audible as a stairstep discontinuity. Even if the discrepancy
in a normal lapped frame is only .1dB (usually far below the
threshhold of perception), that's a sudden cliff of 380 steps in a 16
bit sample (when there's a boundary with no lapping).
<h2>I thought Vorbis was gapless</h2>
<p>It is. Vorbis introduces no extra samples at the beginning or end
of a stream, nor does it remove any samples. Gapless encoding
eliminates 99% of the click, pop or outright blown speaker that would
occur if boundaries had gaps or made no effort to align
transitions. However, gapless encoding is not enough to entirely
eliminate stairstep discontinuities all the time for exactly the
reasons described above.
<p>Frame lapping, like Vorbis performs internally during continuous
playback, is necessary to eliminate that last epsilon of trouble.
<h1>Easiest Crosslap</h1>
The easiest way to perform crosslapping in Vorbis is to use the
lapping functions with no other extra effort. These functions behave
identically to when lapping isn't used except to provide
at-least-very-good lapping results. Crosslapping will not introduce
any samples into or remove any samples from the decoded audio; the
only difference is that the transition is lapped. Lapping occurs from
the current PCM position (either in the old stream, or at the position
prior to calling a lapping seek) forward into the next
half-short-block of audio data to be read from the new stream or
position.
<p>Ideally, vorbisfile internally reads an extra frame of audio from
the old stream/position to perform lapping into the new
stream/position. However, automagic crosslapping works properly even
if the old stream/position is at EOF. In this case, the synthetic
post-extrapolation generated by the encoder to pad out the last block
with appropriate data (and avoid encoding a stairstep, which is
inefficient) is used for crosslapping purposes. Although this is
synthetic data, the result is still usually completely unnoticable
even in careful listening (and always preferable to a click or pop).
<p>Vorbisfile will lap between streams of differing numbers of
channels. Any extra channels from the old stream are ignored; playback
of these channels simply ends. Extra channels in the new stream are
lapped from silence. Vorbisfile will also lap between streams links
of differing sample rates. In this case, the sample rates are ignored
(no implicit resampling is done to match playback). It is up to the
application developer to decide if this behavior makes any sense in a
given context; in practical use, these default behaviors perform
sensibly.
<h1>Best Crosslap</h1>
<p>To acheive the best possible crosslapping results, avoid the case
where synthetic extrapolation data is used for crosslapping. That is,
design loops and samples such that a little bit of data is left over
in sample A when seeking to sample B. Normally, the end of sample A
and the beginning of B would overlap exactly; this allows
crosslapping to perform exactly as it would within vorbis when
stitching audio frames together into continuous decoded audio.
<p>The optimal amount of overlap is half a short-block, and this
varies by compression mode. Each encoder will vary in exact block
size selection; for vorbis 1.0, for -q0 through -q10 and 44kHz or
greater, a half-short block is 64 samples.
<br><br>
<hr noshade>
<table border=0 width=100%>
<tr valign=top>
<td><p class=tiny>copyright &copy; 2000-2010 Xiph.Org</p></td>
<td align=right><p class=tiny><a href="http://www.xiph.org/ogg/vorbis/">Ogg Vorbis</a></p></td>
</tr><tr>
<td><p class=tiny>Vorbisfile documentation</p></td>
<td align=right><p class=tiny>vorbisfile version 1.3.2 - 20101101</p></td>
</tr>
</table>
</body>
</html>