DTS & Data Compression Measurements
The DTS demonstration coder/decoder (codec) is a slim rack-mount unit with both analog and AES/EBU digital input and output jacks. A menu system allows the user to select inputs and outputs, data reduction or bypass, and the effective data rate—706 kilobits/second/channel (the same as the CD standard), 240kb/s/ch (the proposed rate for a 5.1-channel laserdisc system using the PCM stereo tracks), or 128kb/s. A label on the rear proclaims that it was made by Algo Rhythmic Technology of Northern Ireland.
As well as auditioning the unit—I basically agree with TJN that, with the channel bandwidth set to 240kb/s, the unit is virtually transparent, lacking only perhaps a little pace and top-octave air—I carried out a set of spectral analyses using one of the multitone test signals (footnote 1) on Stereophile's Test CD 3. Fig.1 shows the spectrum of the Track 25 test signal, a demanding mix of 43 sinewaves spaced 500Hz apart with the complete signal having a level of -10dBFS. Note that in between the tones the noise floor is below the -120dBFS limit of the graph.
Fig.1 Spectrum of equal-amplitude, 43-tone test signal, with a combined amplitude of -10dBFS (linear frequency scale, 20dB/vertical div.).
Fig.2 shows the effect of the DTS Zeta algorithm set to 240kb/s/channel. All the tones are reproduced at the correct level, but the noise floor rises. Though the noise/quantizing floor now reaches above -80dB at the top of the audio band, below 10kHz the noise still lies below -100dB, suggesting that the effective resolution is at least around 16 bits. The noise floor also reaches its lowest level in the 2-4kHz region, where the ear is most sensitive, implying a subjective resolution in excess of 16 bits.
Fig.2 DTS Zeta demonstration codec, 240kb/s/ch, spectrum of fig.1 signal after encoding/decoding (linear frequency scale, 20dB/vertical div.).
Fig.3 shows the effect of reducing the data rate to 128kb/s. The noise floor is now between 20dB and 30dB higher than in fig.2, and is not nearly so weighted toward the inaudible high frequencies. In addition, although it's hard to see at the scale these graphs are reproduced in the magazine, the amplitudes of the tones are not always preserved, those at higher frequencies tending to have a slight positive error. The very topmost components now also have a severe rolloff. Not surprisingly, I found the effect of the Zeta algorithm set to this lower data rate to be quite audible on music, and not acceptable in a high-end context.
Fig.3 DTS Zeta demonstration codec, 128kb/s/ch, spectrum of fig.1 signal after encoding/decoding (linear frequency scale, 20dB/vertical div.)
One way to magnify the errors introduced by a perceptual coding algorithm is to pass a signal through several devices in series. With just the one box, I synthesized this situation by storing the results of one pass through the device on DAT, then playing that DAT through the box and storing the result on a second DAT, and so on. Fig.4 shows the result on the 500Hz-spaced test signal by passing it five times through the DTS codec. Compared with fig.2, the noise floor has risen by around 10dB, but the overall level of spuriae is still quite low.
Fig.4 DTS Zeta demonstration codec, 240kb/s/ch, spectrum of fig.1 signal after encoding/decoding five times (linear frequency scale, 20dB/vertical div.).
There was one aspect of the cascaded codec tests that particularly bothered me: the gain errors mentioned earlier with the effective data rate reduced to 128kb/s resulted in digital clipping with music samples that originally reached 0dBFS. This will invalidate listening comparisons using such tracks.
How good overall is this measured performance? For reference, figs.5 and 6 show the effect of the two other perceptual coding algorithms Stereophile has tested on the same multitone signal: DCC's PASC and MD's ATRAC. (The absolute level of the signal was a little different in these tests than for the Zeta, but what's important is the shape of the noise floor, the distribution of the noise/quantizing distortion.)
Fig.5 Sony MDS-501, spectrum of fig.1 signal after ATRAC encoding/decoding (linear frequency scale, 20dB/vertical div.).
Fig.6 Marantz DD-92, spectrum of fig.1 signal after PASC encoding/decoding (linear frequency scale, 20dB/vertical div.).
Fig.5 (ATRAC) shows that the 142kb/s MiniDisc encoding both raises the noise floor way above that of either of the DTS Zeta examples—240kb/s and 128kb/s—and discards the signal components above 18.5kHz. In fig.6, it can be seen that, while the level of noise/quantizing spuriae introduced by the 192kb/s PASC algorithm is both lower than ATRAC and psychoacoustically shaped—in that the highest level of error lies in the top octave, where the ear is not very sensitive, and the smallest error in the midrange—it eliminates some of the highest-frequency tones. Again, the DTS Zeta algorithm appears to preserve more of the original signal resolution.
Remember, however, that both the DCC and MD measurements were made on consumer machines where the mathematical accuracy possible will have been compromised by the need to make the products affordable. The DTS box need suffer no such compromise and uses, I assume, cost-no-object digital signal processing. It's possible that the measured differences between ATRAC or PASC codecs executed with the same attention to processor quality as DTS's Zeta would be considerably smaller. Nevertheless, Zeta encoding/decoding appears to be an impressively transparent process, though it remains to be seen how much of its performance will be affected by the need to use affordable decoder circuitry in a consumer laserdisc player.
One final graph: fig.7 shows a similar spectral analysis of a DTS-processed signal, but this time with gaps present where there are no signal frequencies present in the original. Note how the noise floor quickly drops down to the -120dBFS level when there is no signal present to mask spuriae. This is excellent codec performance.
Fig.7 DTS Zeta demonstration codec, 240kb/s/ch, spectrum of multitone signal with no components between 2.5kHz-4.5kHz and 9kHz-12kHz (linear frequency scale, 20dB/vertical div.).
As TJN mentions above, we haven't yet been able to obtain a similar demonstration codec from Dolby, compounded by the fact that the AC-3 algorithm is intended from the outset to be a multichannel device, meaning that just feeding it with AES/EBU two-channel data will not give a true representation of its performance. By comparison, the DTS system's five data channels are each identical to one of the channels assessed here. Nevertheless, I hope we will be able to perform some careful listening tests on AC-3 technology within the near future, as well as performing an identical set of measurements.—John Atkinson
Footnote 1: Although Audio Precision now offers a complete set of specialized tests for assessing the performance of perceptual coders, these didn't arrive in Santa Fe until after the DTS box had been returned.—John Atkinson