<!--#include virtual="/ssi/header.include" --> <title>The CELT ultra-low delay audio codec: CELT 0.6.0 automated testing results</title> <style type="text/css"> <!-- #navlink_ a { text-decoration: underline !important; } --> </style> </head> <body> <!--#include virtual="/common/xiphbar.include" --> <!--#include virtual="/ssi/pagetop.include" --> <h1>CELT 0.6.0 automated testing results</h1> <p> The automated testing routine for CELT involves running roughly 12 months of audio through the CELT encoder and decoder across a wide variety of modes and configurations. All user accessible modes receive at least some level of coverage. 48kHz mono receives automated quality testing of all frame sizes and most reasonable bit-rates. Common configurations receive extensive fuzz testing under valgrind. ARM (OpenMoko), PPC (Fedora 11), x86_64 (Fedora 10, 11), and x86 (Fedora 11) are used in testing. </p> <p> This level of extensive testing is made possible by the large multiple of real-time that CELT operates at on modern computing hardware. </p> <p> Keep in mind that as of 0.6.0 CELT is still a work in progress. Neither the API/ABI, nor the bit-stream are stable. Also, while we do not expect it to set your house on fire, we cannot guarantee that it won't. Spontaneous combustion is specifically not covered by these tests. </p> <h2>Automated Quality testing</h2> <table align="right" border=1> <tr><th>Value</th><th>Meaning</th></tr> <tr><td>0</td><td>Imperceptible</td></tr> <tr><td>-1</td><td>Perceptible but not annoying</td></tr> <tr><td>-2</td><td>Slightly annoying</td></tr> <tr><td>-3</td><td>Annoying</td></tr> <tr><td>-4</td><td>Very annoying</td></tr> <tr><td colspan="2"><center><i>Definitions of PEAQ ODG scores</i></center></td></tr> </table> <p> The quality of CELT 0.6.0 at 48 kHz mono was assessed for 51,848 combinations of bitrate, frame size, and complexity using <a href="http://www-mmsp.ece.mcgill.ca/Documents/Software/Packages/AFsp/PQevalAudio.html">PQEvalAudio</a>, an implementation of <a href="http://en.wikipedia.org/wiki/PEAQ">PEAQ</a>. The PEAQ objective difference grade does not always accurately reflect human opinion but its automated nature permits testing large numbers of configurations. These quality tests would require over 52 days of continuous listening if conducted with a single human reviewer. </p> <br style="clear:both;"/> <h3>Complexity 9 PQEvalAudio map</h3> <p>This illustration demonstrates the quality/bitrate/delay trade-offs available in CELT in full (default) complexity mode. </p> <p><center><table width=500><tr><td><a href="060_9_peaqmap.png"><img src="060_9_peaqmap.thumb.png" border=0 width=500 height=298 alt="CELT 0.6.0 Quality Graph"/></a></td></tr> <tr><td align="right">Equal-quality contours are drawn at -0.5, -1, -2, and -3.</td></tr></table> </center> </p> <h3>Complexity 1 PQEvalAudio map</h3> <p>This illustration demonstrates the quality/bitrate/delay trade-offs available in CELT in low complexity mode. </p> <p><center><table width=500><tr><td><a href="060_1_peaqmap.png"><img src="060_1_peaqmap.thumb.png" border=0 width=500 height=298 alt="CELT 0.6.0 Quality Graph"/></a></td></tr> <tr><td align="right">Equal-quality contours are drawn at -0.5, -1, -2, and -3.</td></tr></table> </center> </p> <h3>Cross-version PQEvalAudio comparisons</h3> <p>PQEvalAudio is run periodically during CELT development to help spot unexpected changes which may be perceptually relevant. Sometimes new functionality will introduce quality impacting bugs which only impact some configurations.<p/> <p>For 0.6.0 this comparative testing appears to show a reduction in performance around 40kbit/sec at typical frame sizes, but this is actually a case of PQEvalAudio disagreeing with actual human listening tests: Towards the end of the 0.6.0 development cycle the codec was retuned based on real listening tests and this tuning is responsible for the decline in the PQEvalAudio score even though the improvement provided by the tuning is obvious to any listener. For comparison graphs are provided comparing 0.6.0 with 9dff0218, a recent version the source code repository immediately prior to these tuning changes. <p/> <p>Several other quality improvements during the 0.6.0 development cycle were partially offset by the introduction of independently coded frames. Now the CELT encoder will automatically encode some frames independently of the prior frames. This makes the stream somewhat more robust to packet loss. Applications can also request the CELT encoder to produce only independent frames, which gives the greatest robustness to packet loss but it requires a somewhat higher bitrate to achieve the same quality.</p> <h4>Comparison with CELT 0.5.2 (complexity 9)</h4> <p>For each test point the 0.5.2 PQEvalAudio score was subtracted from the CELT 0.6.0 score. <p><center><table width=500><tr><td><a href="060-vs-052_9_peaqmap.png"><img src="060-vs-052_9_peaqmap.thumb.png" border=0 width=500 height=298 alt="CELT 0.6.0 Quality Graph"/></a></td></tr> <tr><td align="right"><i>Positive (blue) values in the chart indicate improvement according to PQEvalAudio, while negative (red) values indicate quality loss.</i></td></tr></table> </center> <h4>Comparison with CELT 9dff0218 (complexity 9)</h4> <p>For each test point the revision 9dff0218 PQEvalAudio score was subtracted from the CELT 0.6.0 score. <p><center><table width=500><tr><td><a href="060-vs-9dff0218_1_peaqmap.png"><img src="060-vs-9dff0218_1_peaqmap.thumb.png" border=0 width=500 height=298 alt="CELT 0.6.0 Quality Graph"/></a></td></tr> <tr><td align="right"><i>Positive (blue) values in the chart indicate improvement according to PQEvalAudio, while negative (red) values indicate quality loss.</i></td></tr></table> </center> <h4>Comparison with CELT 0.5.2 (complexity 1)</h4> <p>For each test point the 0.5.2 PQEvalAudio score was subtracted from the CELT 0.6.0 score. <p><center><table width=500><tr><td><a href="060-vs-052_9_peaqmap.png"><img src="060-vs-052_9_peaqmap.thumb.png" border=0 width=500 height=294 alt="CELT 0.6.0 Quality Graph"/></a></td></tr> <tr><td align="right"><i>Positive (blue) values in the chart indicate improvement according to PQEvalAudio, while negative (red) values indicate quality loss.</i></td></tr></table> </center> <h4>Comparison with CELT 9dff0218 (complexity 1)</h4> <p>For each test point the revision 9dff0218 PQEvalAudio score was subtracted from the CELT 0.6.0 score. <p><center><table width=500><tr><td><a href="060-vs-9dff0218_1_peaqmap.png"><img src="060-vs-9dff0218_1_peaqmap.thumb.png" border=0 width=500 height=298 alt="CELT 0.6.0 Quality Graph"/></a></td></tr> <tr><td align="right"><i>Positive (blue) values in the chart indicate improvement according to PQEvalAudio, while negative (red) values indicate quality loss.</i></td></tr></table> </center> </p> <h2>"make check" tests</h2> <p>CELT includes a number of unit tests that exercises internal components of CELT.<br/> 0.6.0 introduces a new test 'tandem-test' which loops the decoder output back into the encoder and tests the whole encoder/decoder system at many rates and frame sizes.</p> <table> <tr><th>Test</th><th>x86_64</th><th>x86</th><th>ARM</th><th>PPC</th></tr> <tr><td>cwrs32-test</td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td></tr> <tr><td>dft-test</td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td></tr> <tr><td>ectest</td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td></tr> <tr><td>laplace-test</td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td></tr> <tr><td>mathops-test</td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td></tr> <tr><td>mdct-test</td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td></tr> <tr><td>real-fft-test</td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td></tr> <tr><td>tandem-test</td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td></tr> <tr><td>type-test</td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td><td><font color="green"><b>Pass</b></font></td></tr> </table> <h2>All modes test</h2> A short audio file is run through 27,525,120 CELT configurations (all frame sizes, all bytes-per-frame from 8-200, and sample rates from 32000-96000 in 100Hz increments). Because of CPU requirements this test is only run only in low complexity mode. In order to pass, these cycles of "testcelt" must complete without error. <ul> <li> x86_64: <font color="green"><b>Pass</b></font> <li> x86_64 fixed point: <font color="green"><b>Pass</b></font> <li> x86: <font color="green"><b>Pass</b></font> <li> x86 fixed point: <font color="green"><b>Pass</b></font> <li> PPC: <font color="green"><b>Pass</b></font> <li> PPC fixed point: <font color="green"><b>Pass</b></font> </ul> <h2>Popular modes fuzz-test</h2> Two hours of audio extracted from several dozen albums and live recordings are run through CELT at 32, 44.1, and 48 kHz at frame sizes of 64, 96, 128, 192, 256, 384 and 512 samples and at 48, 64, and 128kbit/sec in mono, stereo mode, and with and without VBR. One tenth of a percent of the encoded bits are randomly flipped. In order to pass, these cycles of "testcelt" must complete without error. This test is run under valgrind both with the memtest and exp-ptrcheck tools and with assertions enabled for extra error sensitivity. <ul> <li> x86_64: <font color="green"><b>Pass</b></font> <li> x86_64 alloca (psedo-stack mode): <font color="green"><b>Pass</b></font> <li> x86: <font color="green"><b>Pass</b></font> <li> x86 fixed point: <font color="green"><b>Pass</b></font> <li> PPC: <font color="green"><b>Pass</b></font> <li> PPC fixed point: <font color="green"><b>Pass</b></font> </ul> <!--#include virtual="/ssi/pagebottom.include" -->