Mirror of opus
21af73eb21
Jean-Marc's original anti-collapse patch used a threshold on the
content of a decoded band to determine whether or not it should
be filled with random noise.
Since this is highly sensitive to the accuracy of the
implementation, it could lead to significant decoder output
differences even if decoding error up to that point was relatively
small.
This patch detects collapsed bands from the output of the vector
quantizer, using exact integer arithmetic.
It makes two simplifying assumptions:
a) If either input to haar1() is non-zero during TF resolution
adjustments, then the output will be non-zero.
b) If the content of a block is non-zero in any of the bands that
are used for folding, then the folded output will be non-zero.
b) in particular is likely to be false when SPREAD_NONE is used.
It also ignores the case where mid and side are orthogonal in
stereo_merge, but this is relatively unlikely.
This misses just over 3% of the cases that Jean-Marc's anti-collapse
detection strategy would catch, but does not mis-classify any (all
detected collapses are true collapses).
This patch overloads the "fill" parameter to mark which blocks have
non-zero content for folding.
As a consequence, if a set of blocks on one side of a split has
collapsed, _no_ folding is done: the result would be zero anyway,
except for short blocks with SPREAD_AGGRESSIVE that are split down
to a single block, but a) that means a lot of bits were available
so a collapse is unlikely and b) anti-collapse can fill the block
anyway, if it's used.
This also means that if itheta==0 or itheta==16384, we no longer
fold at all on that side (even with long blocks), since we'd be
multiplying the result by zero anyway.
|
||
|---|---|---|
| doc/ietf | ||
| libcelt | ||
| tests | ||
| tools | ||
| .gitignore | ||
| AUTHORS | ||
| autogen.sh | ||
| celt.kdevelop | ||
| celt.pc.in | ||
| ChangeLog | ||
| configure.ac | ||
| COPYING | ||
| Doxyfile | ||
| Doxyfile.devel | ||
| INSTALL | ||
| libcelt.spec.in | ||
| Makefile.am | ||
| NEWS | ||
| README | ||
| README.Win32 | ||
| TODO | ||
CELT is a very low delay audio codec designed for high-quality communications. Traditional full-bandwidth codecs such as Vorbis and AAC can offer high quality but they require codec delays of hundreds of milliseconds, which makes them unsuitable for real-time interactive applications like tele- conferencing. Speech targeted codecs, such as Speex or G.722, have lower 20-40ms delays but their speech focus and limited sampling rates restricts their quality, especially for music. Additionally, the other mandatory components of a full network audio system— audio interfaces, routers, jitter buffers— each add their own delay. For lower speed networks the time it takes to serialize a packet onto the network cable takes considerable time, and over the long distances the speed of light imposes a significant delay. In teleconferencing— it is important to keep delay low so that the participants can communicate fluidly without talking on top of each other and so that their own voices don't return after a round trip as an annoying echo. For network music performance— research has show that the total one way delay must be kept under 25ms to avoid degrading the musicians performance. Since many of the sources of delay in a complete system are outside of the user's control (such as the speed of light) it is often only possible to reduce the total delay by reducing the codec delay. Low delay has traditionally been considered a challenging area in audio codec design, because as a codec is forced to work on the smaller chunks of audio required for low delay it has access to less redundancy and less perceptual information which it can use to reduce the size of the transmitted audio. CELT is designed to bridge the gap between "music" and "speech" codecs, permitting new very high quality teleconferencing applications, and to go further, permitting latencies much lower than speech codecs normally provide to enable applications such as remote musical collaboration even over long distances. In keeping with the Xiph.Org mission— CELT is also designed to accomplish this without copyright or patent encumbrance. Only by keeping the formats that drive our Internet communication free and unencumbered can we maximize innovation, collaboration, and interoperability. Fortunately, CELT is ahead of the adoption curve in its target application space, so there should be no reason for someone who needs what CELT provides to go with a proprietary codec. CELT has been tested on x86, x86_64, ARM, and the TI C55x DSPs, and should be portable to any platform with a working C compiler and on the order of 100 MIPS of processing power. The code is still in early stage, so it may be broken from time to time, and the bit-stream is not frozen yet, so it is different from one version to another. Oh, and don't complain if it sets your house on fire. Complaints and accolades can be directed to the CELT mailing list: http://lists.xiph.org/mailman/listinfo/celt-dev/ To compile: % ./configure % make For platforms without fast floating point support (such as ARM) use the --enable-fixed argument to configure to build a fixed-point version of CELT. There are Ogg-based encode/decode tools in tools/. These are quite similar to the speexenc/speexdec tools. Use the --help option for details. There is also a basic tool for testing the encoder and decoder called "testcelt" located in libcelt/: % testcelt <rate> <channels> <frame size> <bytes per packet> input.sw output.sw where input.sw is a 16-bit (machine endian) audio file sampled at 32000 Hz to 96000 Hz. The output file is already decompressed. For example, for a 44.1 kHz mono stream at ~64kbit/sec and with 256 sample frames: % testcelt 44100 1 256 46 intput.sw output.sw Since 44100/256*46*8 = 63393.74 bits/sec. All even frame sizes from 64 to 512 are currently supported, although power-of-two sizes are recommended and most CELT development is done using a size of 256. The delay imposed by CELT is 1.25x - 1.5x the frame duration depending on the frame size and some details of CELT's internal operation. For 256 sample frames the delay is 1.5x or 384 samples, so the total codec delay in the above example is 8.70ms (1000/(44100/384)).