The 1st Test
With one to three minutes apiece, the total test time of 14 minutes for the original audiosignals is not negligible. The procedure was: listen to the original, compare with the mp3 and repeat if uncertain. In addition, the quality of some of the recordings were tested after they were run through the coding process twice. To help judge the degree of mp3 distortion, three of the original CD recordings were also played as WAV files over a simple PCI soundcard (Ensoniq by Creative) across a high quality 12m cable running from the headphone output of the CD player. The wave editor was used subsequently to adust the volume to that of the CD originals, since even small volume differences can affect a hearing test.
Even though it has a certain influence on the evaluation of the sound test: The synthetic signals were evaluated first using CoolEdit since they did not require an elaborate hearing test. The pulse signal is quite apparently distorted, especially since ringing of a magnitude of several percent is obvious. Interestingly enough, both channels of the monosignal were processed significantly differently. A subsequent mp3 compression resulted in practically no change in the signal form. Analog amplifiers were considerably better in this respect, speakers considerably worse. The differences in the results from the other two decoders were relatively small (peaking and noise of the same magnitude). Mpeg Encoder V.0.06 was clearly superior for mp3 coding, however (no peaking, noise of the same magnitude), but also encoded much slower (ca. 25 minutes for a minute and a half of music on a 233MHz pc!). One might maintain that such synthetic signals have nothing to do with music, but that argument doesn’t hold in the age of techno-pop.
The evaluation of the two-tone signal was also very interesting. Additional discrete spectral lines would result in analog circuits due to the nonlinearities, e.g. sum and difference frequencies. For mp3 though, the signal only displays minimal noise in the time domain. A frequency analysis per FFT brings the facts to light. No new discrete distortion lines are introduced, but a noise veil is generated about the original signals. RF engineers recognize such noise formation, e.g. from the phase noise in oscillators. The double conversion did not have significant influence in this case either.
Figure 1 : mp3 behavior for a sinus burst. Some encoders result
in less peaking, but the noise veils are typical. Only the LAME
encoder was nearly perfect at 128kBit/s, others needed twice the
rate!
Figure 2 : mp3 distortion in a two-tone spectrum. The original is
below. Onlx the LAME was nearly perfect, others not - even at
256kBit/s.
After the initial success of this detective work, the Beethoven recording was closely examined on account of the many impulses it contains. The wave editor shows nice pulses here as well, but in this case they remain optically distinct. This is fairly easy to explain, since any over- or undershoot is mixed in with the following tones. More exact information can be obtained by simply taking the differential signal, which should be zero in the absence of conversion error. It is not that simple in practice, however, since even a simple delay time would result in a nonzero difference. And in fact, the results show just that. Figure 3 clearly shows a time shift of ca. 23 ms.
Figure 3 : mp3 above and the original below, the left channel in
both cases
Figure 4 : Detailed view: mp3 above and the original below, after
compensating for the transmission delay
If the transmission delay is compensated, the signals which remain in the time and frequency domain are far less conspicuous to the eye than those for the synthetic signals. There is also no trace of noise veils. Only when the difference is taken following delay compensation is it possible to recognize that the eye can be fooled and the differences are substantial. This does not have to result purely from nonlinear distortion however. Effects such as, e.g. group delay variations over frequency could also contribute. Mp3 therefore wins this round. Repeating the procedure using analog recording shows that better results are attainable (Figure 5).
Figure 5 : Which is the original? CD vs. Analog recording (at
bottom) for "Without You" by Mariah Carey
But we won’t be satisfied that easily and have therefore chosen to superimpose an number of pulses on the quiet start of "Without You" by Mariah Carey. The burst level was held to -26dB (0dB=full scale drive) to avoid overdriving and thus was detectable only as a mild peeping. The recording was converted to mp3 and the converted bursts were subtracted. Just as in the WAV-format, the burst was as good as gone, although minimal residual could be seen in the wave editor. No disturbances were detectable per soundcard or headphones. The same test was then repeated with a -9dB pulse level (still significantly lower than the clipping level). Audible and visible effects became apparent for this case. The peaks in Figure 6 have the same effects as scratches on old vinyl records. Following these prep tests, the question remained: How does the ear react to mp3 for normal music?
Figure 6 : mp3-Signal with burst increased
The Soundtest
Detective Needed
Conclusion
The Audio Equipment
T&A TMR 1 speakers
Onkyo amp Integra A8780
Sennheiser headphones HD 580 precision
CD-Player Sony X33 ES
Ensoniq soundcard from Creative in a PII-PC with 233MHz
TEAC CD-burner (recorded at 2x)
Overall lower high-quality with good price to performance ratio.
Software Used
Encoders : CDex Version 1.10 (for comparison: mpeg Encoder
V.0.06, NexEncode 2.0, Sound Limit 2.5, Ultimate Encoder 1.05,
MpegDJ Encoder 1.97, LAME 3.87, MusicMatch 5.10 and FHG-Test
Version)
Decoders : Dema mp3-Softwareplayer Version 2.0.0 (for comparison:
Audioactive MP3 Decoder V1.1.2 and CoolDecode 1.4)
For analysis and signal generation : CoolEdit 96 For CD burning :
WinOnCD 3.5
All software is available via the Internet. The actual coding done by the encoders is typically done using standard modules and only the interace is actually original. Thus the eight programs actually contain only 5 different encoders.
| To Top - To Homepage - German Version |
 |
Copyright © 2000 Stephan Weber. All rights reserved.
Stand: März 03, 2001.