Nieuws & Weblog

Audio Myth - "DSD: a direct stream from A/D to D/A."

  Monday 31 August 2015, 10:56
  Weblog
This myth goes something like this:

"DSD provides a simple and direct digital path between the A/D and D/A."
"DSD is simpler than PCM."
"DSD is not PCM."
Audio Myth - "DSD: a direct stream from A/D to D/A."

While DSD can provide spectacular audio performance, all of the statements above are false. DSD is 1-bit PCM, and in almost all practical systems, it greatly complicates the path between the A/D and D/A converter. The name "Direct Stream Digital" or DSD is intended to imply that the 1-bit path is a simpler more direct path that requires less digital processing. This is completely misleading. There are only a very few specialized applications where DSD can create a simple processing path. In commercial DSD releases, the direct path is a myth.

There are many spectacular DSD recordings, but the quality is not due to any virtues of the DSD format.

Measured Performance
DSD is a high-resolution digital audio format, and few recordings can fully utilize the SNR and bandwidth available in the DSD format. Nevertheless, the measured performance of DSD falls between that of the CD and 96 kHz 24-bit PCM. DSD can achieve a 120 dB 20 Hz to 20 kHz SNR with a usable bandwidth of about 50 kHz. For this reason, DSD is almost exactly equivalent to a 20-bit 96 kHz PCM system. Notice that I said "20-bit" and not "24-bit". In theory, a 24-bit system is about 24 dB quieter than DSD, but it is virtually impossible to use all of the available SNR in either system.

A Simple and Attractive Concept
DSD seems like a simple and attractive concept. In theory, the output of a 2.8 MHz 1-bit oversampled A/D converter could be wired directly to the input of a 1-bit oversampled D/A converter. With this direct connection it would not be necessary to convert the 2.8 MHz signal to and from a multi-bit signal at a lower sample rate. In 1995 this seemed to make sense to Sony. At that time, most converters had an internal 1-bit 2.8224 MHz conversion system (64 x the CD sample rate). Sony proposed direct storage and retrieval of this 2.8224 MHz "bit stream" for archival purposes.

Archival Applications
Given the 1-bit architecture of converter chips in 1995, the Sony bit-stream archival system provided the simplest path for storage and retrieval of completed recordings. Their goal was to digitally record analog tapes, vinyl records, and historic recordings using the best digital techniques available. The bit-stream archival system bypassed the 1-bit to 16-bit to 1-bit (or 1-bit to 24-bit to 1-bit) processing of typical 1995 vintage digital systems. In this limited context, the bit-stream system was a simpler and cleaner digital path.

By 1995 a few 24-bit 88.2 kHz PCM systems had been introduced. This led to disagreements over which word length and sample rate should be used for archival purposes. Sony's bit-stream archival system skirted this issue by providing a system that could easily be converted to 44.1 or 88.2 kHz. This was a great solution until oversampled converters progressed beyond 1-bit technology.

The Death of 1-bit Conversion
All 1-bit converters produce high levels of ultrasonic noise and they tend to have troublesome low-level idle tones within the audio band. Each additional bit reduces the total noise output by 6 dB. In the late 1990's, this prompted a move to 1.5, 2, and 3-bit oversampled conversion, and the 1-bit converters became obsolete.

Most modern oversampled multi-bit D/A converters use a number of equally weighted 1-bit converters running in parallel. One early example is Benchmark's 1997 introduction of the DAC2004, a 20-bit D/A converter built with a 2-bit internal architecture. At that time, we used two one-bit Phillips TDA1547 "bit-stream" converters to create the 2-bit architecture. A short time later, IC manufacturers began introducing single-chip solutions.

The best converters available today use 4-bit to 6-bit internal architectures. For example, our DAC2 HGC uses eight 6-bit converters running in parallel at a very high internal sample rate. Each of these 6-bit converters are constructed from an array of 64 1-bit switched current sources running in parallel. This means that each channel of the DAC2 HGC has 512 1-bit converters running in parallel. When compared to 1-bit systems, these massively parallel architectures achieve much lower noise and distortion, while virtually eliminating ultrasonic noise.

The Near-Death of the 1-bit System
The Sony bit-stream archival system failed to gain much traction, and 1-bit systems would have disappeared entirely if it had not been for the SACD. Sony and Phillips had jointly developed the CD and this had generated significant royalty income. In 1999, Sony and Philips joined forces again to introduce the SACD (Super Audio CD) as a successor to the CD. This launched a format war between the new audio-only DVD-Audio disk and the SACD. Sony launched an aggressive marketing campaign to convince the public that their 1-bit system was superior to the 24-bit formats used on the DVD-Audio disk. In the process, they renamed the bitstream system. The new name, "Direct Stream Digital" is better known as DSD. The DVD-A vs. SACD format war was a repeat of the VHS vs. Betamax, but this time both parties lost. By 2007 the DVD-A and the SACD had both failed to gain any significant market share. Nevertheless, many very high-quality recordings had been released. Fortunately many of these fine recordings are now being made available as downloads in DSD and high-resolution PCM formats.

In the process of aggressively promoting the failing SACD format, many questionable claims were made. Most claims tried to leverage the appeal of the apparent simplicity of the 1-bit system. Other claims highlighted the bandwidth of the 2.8 MHz sampling system while ignoring the fact that a 50 kHz lowpass filter was required at the output of the D/A converter. These claims produced a loyal DSD following. Many SACD releases were excellent recordings, but their quality has nothing to do with any virtues of the DSD format.

Direct Stream Digital is not "Direct"
It is nearly impossible to mix and edit in the 1-bit DSD environment. For this reason, there are almost no DSD recordings that are the "direct" output of an A/D converter. Many DSD releases were produced using high-resolution PCM editing and mixing equipment. Others were produced using analog mixing through several generations of A/D and D/A conversion. Some were mixed to analog tape. In short, there is nothing "direct" in most "Direct Stream Digital" recordings. "Direct" is a myth.

Common DSD Production Methods
The most common production signal paths for creating a DSD release are as follows:
    24-bit PCM A/D > PCM Workstation > 24-bit PCM > 1-bit Noise Shaper > DSD
    1-bit DSD A/D > 1-bit DSD D/A > Analog Console > 1-bit A/D >DSD
    Analog Tape > Analog Mixing > Analog Master > 1-bit A/D > DSD
    1-bit DSD A/D > DSD to DXD Conversion > DXD PCM Mixing > 1-bit Noise Shaper > DSD

The Mythical DSD Production Method
The direct path from a DSD A/D looks like this:
    1-bit DSD A/D > DSD

This direct path offers no possibility of mixing or editing. It is a total myth. With one or two rare exceptions this direct path has never been used for commercial DSD releases. It can provide a direct link between a 1-bit A/D and a 1-bit D/A, but it cannot fully leverage the technological benefits provided by the parallel array of 1-bit converters used in modern multibit converters.

DSD Production Method 1 - 24-bit PCM to DSD Conversion
If we chose to create a DSD recording using conventional multibit digital mixing and editing, a 24-bit to 1-bit conversion will be required for our DSD release. This multibit mixing and editing process is virtually noise free, but the final 24-bit to 1-bit DSD conversion adds noise and distortion to the DSD release. It would be better to simply release the 24-bit version. The DSD release is clearly less "direct" than the 24-bit version. The DSD release will have all of the defects of the 24-bit system plus all of the defects of the 1-bit system.

The culprit is the 1-bit noise-shaping process that is necessary to create any usable 1-bit audio stream. The 1-bit digital system creates massive digital quantization errors which must be shifted or "noise shaped" into the ultrasonic band above 50 kHz. This process leaves the 20 to 20 kHz region of 1st generation DSD with the equivalent of a 20-bit (120 dB) SNR.

If the audio exists in a multibit format, the quality will not be improved by noise-shaping down to 1-bit. This multibit to 1-bit conversion should be avoided when possible.

DSD Production Method 2 - Analog Mixing
The multibit PCM system can be eliminated form the signal chain if an analog mixer is used instead of a digital mixer. This requires an extra DSD D/A and A/D conversion to get in and out of the analog console. Each DSD A/D conversion process requires a 1-bit noise-shaping process. We now have two cascaded 1-bit noise shapers in our signal chain! To make matters worse, each DSD D/A conversion requires a 50 kHz low-pass filter. We now have two cascaded 50 kHz low-pass filters in our signal chain. Furthermore, the analog console will add its own noise and distortion. Again, this production path fails to meet the "direct" path promised by DSD. The measured performance of this path is not nearly as good as that of method 1 above.

DSD Production Method 3 - Analog Tape
This method is a step back to obsolete analog tape technology. All of the defects of analog tape and analog mixing are now part of the chain. We have eliminated cascaded 1-bit noise shapers, cascaded low-pass filters, and cascaded A/D and D/A converters but we have added at least one or two generations of analog tape. This is not a good trade-off and the quality will suffer. This is not a "direct" path.

DSD Production Method 4 - DXD
The desire to avoid the use of conventional PCM (method 1), as well as the obvious performance problems of methods 2 and 3, led to the development of Digital eXtreme Definition (DXD). DXD is a 24-bit 352.8 kHz PCM system. This system has a 144 dB SNR and a 176.4 kHz bandwidth. However, this "extreme" resolution offers no practical advantage over a 24-bit 96 kHz system. The frequency region between 48 kHz and 176.4 kHz will have no audio information and it is almost entirely above the 50 kHz upper limit of the 64X DSD lowpass filter. Furthermore the 24-bit to 1 bit noise shaping used in method 1 above is required when converting DXD to DSD. This method also fails to deliver a direct path between the A/D and the D/A. The bottom line is that this DXD method will provide exactly the same measured performance as that of method 1 above.

Comparison of DSD Production Methods
DSD production methods 1 and 4 will provide the highest performance, but in both cases, the final conversion to DSD is an unnecessary extra processing step. It would be better to deliver the 24-bit product to the consumer. Method 3 will deliver the poorest measured performance, and method 2 will be slightly better than method 3. None will match the measured performance of a conventional 24-bit 96 kHz production and delivery system.

Conclusions
In virtually all cases, DSD adds significant processing to the music production chain. DSD absolutely fails to deliver a "direct" path between the A/D and the D/A. Any such claims are marketing spin. DSD offers no advantages over a conventional 24-bit 96 kHz system, and it fails to leverage the massively parallel architecture of modern converters. DSD cannot match the measured performance of a conventional 24-bit 96 kHz system. Conventional PCM systems provide the most direct and transparent signal path between the A/D in the studio and the D/A in the consumer's home. For these reasons, there is no compelling reason to pay extra for a DSD recording if a 96 kHz version is available. If the choice is between a CD and a DSD version, the DSD version may offer some improvement.

Benchmark recognizes that there are many fine high-resolution recordings that are only available in DSD format. For this reason, Benchmark DAC2 converters are designed to directly accept 24-bit PCM or 1-bit DSD without adding any internal format conversions. This versatility makes it easy to play both high-resolution formats to their fullest potential.

John Siau
Author