It is a well-publicized trend that HDTV (high-definition TV) is sweeping through the next generation of SOC (system-on-chip) platforms. SOCs for set-top boxes, television monitors, disk players, and—in the near future—mobile media players will have HD capability. But another aspect of this evolution, audio quality, is slipping beneath the radar.
The problem with high-resolution audio isn’t just in the circuit design. In fact, analog-IC designers are producing DACs and amplifiers that are arguably better than anything available during the golden age of discrete-component gear. The problem is in characterization and testing. The quality of high-end audio, as many analog designers will tell you and as audiophiles may insist to the point of religious discourse, is excruciatingly hard to quantify even on the characterization bench and almost impossible to verify in a manufacturing-test environment. SOC designers are now sizing up this new challenge, joining experienced audio-IC designers who have lived with it.
So what’s the problem?
The new test-and-measurement challenge is coming from the convergence of two forces. One, as previously mentioned, is the increasing capacity for digital bit streams to provide quality audio. To be more precise, the increasing quality of the source material is not the problem. One can always convert a good bit stream into a mediocre audio signal. The problem is the growing expectations of consumers. When digital audio means audio from an MP3 bit stream or a similarly lossy source, the underlying issues in the codec usually make it unnecessary to take care with the analog circuitry. Users judged MP3 players against portable cassette players and CD players, which for the most part sounded worse in comparison.
“The quality of outputs for MP3 players has actually been driven by the emergence of high-performance headphones rather than by the source,” says Gary Adrig, director of marketing for audio products at National Semiconductor. “As headphones improved, we saw some customers in what had been an undemanding application start to ask for 100-dB SNRs [signal-to-noise ratios] and 0.05% THD [total harmonic distortion].”
As content vendors began to move to lower compression ratios—hence, higher bit rates—chip designers had to move to wider datapaths and better DACs to keep the noise floor of the hardware below the inherent noise level of the decoded source. Despite the limitations of MP3, market competition proved that consumers are quite discriminating about sound quality.
With the new HD media, the pace has picked up as audio tracks have jumped right past CD quality of 16 bits at 44.1k samples/sec to as much as 24-bit, 192k-samples/sec data for DVD Audio. This performance will certainly lead buyers of high-end gear to question the quality of the sound they are hearing. “We are already there in the home-theater market,” says Texas Instruments Marketing Manager Kevin Belnap. “Once you reach a minimum level of noise and harmonic distortion, a lot of listener-preference issues, like sound stage and the whole 'vacuum-tube-sound’ thing, come into play.”
But the critical ear won’t stop there. Users of set-top and converter boxes, HD-ready television sets, and even mobile devices are likely to also demand much better audio than they have been hearing. The course of this evolution might be more apparent from one professional-audio developer’s experience. Morten Lave, chief executive officer of TC Applied Technologies, a developer of studio-monitor speakers, relates his experience with MP3. “I had an iPod, so I decided to put some music on it,” he says. “I used the default settings, and the result was bad; the cymbals sounded awful. So I jacked the bit rate up to 192 kHz. That got me quality as good as the cheap headphones on the iPod, but, if I hook it up to my home system, I can still easily hear compression artifacts.” With the new lossless-data types, the electronics will no longer be able to hide behind the limitations of MP3 compression. They will themselves become the issue.
This escalating demand for audio quality by itself is manageable. Most SOC vendors in the set-top-box and TV markets, in which space and cost are not draconian overlords, now simply provide a digital output to an external analog chip or chip set. This approach passes the characterization and test problems on to companies in the analog world, which is familiar with them. But the second of two converging forces is closing this loophole for SOC-design teams. That force is integration.
As one SOC vendor puts it, the pervasive demand for greater integration across this market is forcing SOC vendors to put DACs—and later, small power amplifiers—onto the main die. This approach not only dredges up all of the widely discussed problems of precision analog design in a noisy, low-voltage digital-CMOS environment, but also drops the characterization and test problems squarely back in the laps of the SOC team. And it’s not going to rest there gracefully.
“We’ve seen the issues that integration can create,” Belnap says. “The early MP3-player guys tried integrating a pulse-width-modulation processor and a DAC onto their chip, but the quality just wasn’t there. Now, with HD DVD or Blu-Ray getting integrated into home receivers, we’re talking about more challenging integration and a whole new level of sound quality.”
High-quality-audio outputs
You can estimate the problem of characterizing the coming wave of SOCs by comparing the way SOC designers typically characterize analog outputs against the techniques that are emerging at the high end of the audio market. This comparison will result in a rather depressing statement of work for SOC designers. Until recently, SOC characterization has focused on the digital side of the audio problem. Here, standards organizations are more than happy to step in and help, providing source bit streams to stimulate the digital input and references against which to compare the results. In the case of lossy-compression systems, these references are envelopes that define an acceptable range of outputs.
“This started about 10 years ago with Dolby Digital for the ATSC [Advanced Television Systems Committee] and DVD video standards,” says Matthew Watson, software-infrastructure manager at Texas Instruments. “They provided results plots for the Audio Precision test equipment, so you could run THD, SNR, and spectral plots and see if you fit.”
Among the most assertive in providing such characterization support has been Dolby—for its own codec IP (intellectual property), of course, but third-party organizations, such as THX, have also been active, Watson says. For lossless-audio formats such as CDs, there is no need to provide envelopes—characterization engineers can compare the output bit stream to a reference stream, and standards organizations can insist on bit-exact output. “The digital-output quality is so far above what the analog section can do that our job is basically done when we have met the external standards,” Watson says. “Customers understand the rigor of the test procedures, and they accept the results.”
But the situation differs greatly on the far side of the DAC. Most chip architects have avoided analog outputs from SOCs, more because of circuit-design and silicon-area issues than testing issues. When architects integrate analog audio, quality expectations have often been low, and characterization has been somewhat perfunctory—examining the analog output at zero and full-scale digital inputs to verify offset and voltage swing and perhaps looking at one output waveform, for instance.
But that situation is changing. As audio moves toward the rarified atmosphere of the audiophile, characterization becomes not only more rigorous, but also more customer-dependent. “At the high end, everyone has a different perspective on the importance of quality,” says Julian Hayes, vice president of marketing at high-end-audio-chip vendor Wolfson Microelectronics. “This leads you to a proliferation of characterization procedures.” Characterization also becomes harder.
The analog output
Experienced vendors of precision analog break the characterization problem into a number of related questions. What should we measure? How do we make the measurements and under what circumstances? How far do we go? And, in the case of high-end audio, another question looms at the end of the process: Can any amount of measurement give us the right answer? These questions are nontrivial because the goal of characterization is not to determine the electrical performance of the output, but to predict the listening experience. That challenge is far more serious.
Just the question of what to measure turns out to be contentious. For indifferent-quality audio, functional measurements suffice. Frequency response, THD, and some sort of noise measurement are generally enough to determine that an audio section will sound OK to a casual listener through inexpensive headphones. These tests have survived since the early days of high fidelity and are still a starting point. And fortunately for engineers, they all come neatly packaged in a single automated box.
“Everybody has an Audio Precision box these days,” says Philippe Mora, director of marketing and business development at Nvidia’s newly acquired PortalPlayer division. In recent years, Audio Precision has combined signal generation and acquisition with analysis and PC-based control to become a de facto standard in audio characterization. By combining the hardware with precoded scripts for measurement sequences, Audio Precision has been able to automate not only the traditional audio measurements, but also many of the procedures that third-party standards organizations require.
No one questions the Audio Precision system’s ability to deliver accurate measurements, even at the extremes of 24-bit data and 192-kHz sample rates. But some designers also warn that the Audio Precision equipment is only a partial answer. “The AP is our main characterization tool for audio outputs,” says Jeff Bridges, audio-applications director at National Semiconductor. “But for specific tests, we will bring in other measuring equipment, as well—usually off-the-shelf tools, like a network analyzer or a spectrum analyzer.” This approach has a tendency to give the characterization bench a satisfying mad-scientist look (Figure 1). But it can also mean a lot of manual steps in the characterization process.
“The range of characterization procedures we see in the industry now is amazing,” suggests Wolfson Chief Technology Officer Peter Frith. “You see some people setting the input to zero and putting a voltmeter on the output to measure noise and then looking at a full-scale sine wave with a scope to measure dynamic range. Others are more traditional: THD, SNR, and dynamic range. But for our markets, that is just the start.” Characterization must also include system-related issues. Particular among these is supply-noise rejection, which in itself is hard to even define when the analog output is coming from a chip with substantial digital content and many operating modes. But it can be critical to sound quality.
The fact that SOCs are digital devices with analog outputs creates another category of characterization problems. “Early on, there was no industry standard for click and pop noise in the analog output,” says Wolfson’s Hayes. “Similarly, the so-called zipper noise that systems can generate as they step through the levels on a digital gain control was new to the audio world. These noise sources, because they come from specific user actions, would never show up on a traditional characterization. But they can be annoying or even injurious if you have a high-performance headphone stuffed in your ear. So, we had to develop characterization tests for them.”
The unidentifiable
But a serious problem remains. “It can happen that an amplifier measures well and just sounds bad,” admits TI’s Belnap. This statement does not in any way side with what many engineers consider to be a delusional movement among high-end audiophiles—the sort that demagnetizes vinyl records and seeks out hand-braided, gold-plated unobtainium speaker cable. It is rather to admit that the human ear is sufficiently sensitive and adaptable that there is no one quantitative test that can predict how a given DAC, amplifier, and speaker combination will sound to experienced listeners.
This reality has already struck high-end-audio-IC vendors. “At the high-performance end of our market, we used to walk in with sample parts and spec sheets and show customers the data,” relates National Semiconductor’s Bridges. “But recently, more customers have been asking us to leave the spec sheet at home and bring in a working reference design. They take it right into their sound room and start listening to it. At the high end today, quality is all about how the chip sounds.”
This situation presents some obvious problems. For one thing, the fixtures designers use for characterization are often poor representations of a real listening environment. “Almost everyone does their data sheets based on resistive loads,” Bridges observes. But only a very stable amplifier behaves in the same way with a resistive load as it does with a dynamic, reactive load, such as a loudspeaker. In fact, TC Applied Technologies’ Lave suggests that at least for Class D and fully digital amps, the problem of controlling the speaker cone—or, ultimately, the sound-pressure level on the surface of the cone—is sufficiently speaker-dependent that active speakers—those with dedicated built-in amplifiers—will predominate in the industry. It’s simply too hard to build an amplifier that can control all the dynamics that any conceivable speaker network can throw at an output stage.
That problem is not the last one, either. Interacting with skilled listeners often uncovers audible issues that are perfectly measurable—if you know what you are looking for (see sidebar “Can you even measure that?”). Some listener observations that sound totally irrational turn out to be not only reproducible in blind tests, but also traceable back to something that actually does show up in measurements. Each one of these experiences adds its own contribution of complexity to the characterization process.
The result is a characterization process that tends to produce a “sound” that listeners begin to associate with a manufacturer. In some cases, vendors strive to make the sound as dry, or neutral, as possible. “We aim for a dry sound—as much as possible consistent with the price point,” says TI Senior Applications Engineer Fred Shipley. “That allows our customers to manipulate both their digital-signal processing and their board-level analog design to create their own characteristic sound, rather than having to work with ours.” Shipley adds that part of the process of characterizing chips for this TI dry sound is a significant amount of time in the listening room, exposing the chips in TI’s reference design to the company’s golden ears.
When the SOC is the board-level design, the responsibility for the ultimate sound falls on the chip designers. And that sound may be more market-related than specification-based. PortalPlayer’s Mora observes, “You need empirical testing, as well. What’s 'right’ to listeners depends, among other things, on their culture and listening habits. For instance, as a generalization, Asian markets tend to prefer emphasis on the higher frequencies in the spectrum. Europeans tend to think a flatter frequency response but more volume sounds more natural.”
So in the end, is characterization quantitative or qualitative? “I subscribe to both sides of that debate,” says Bruce Hofer, chairman and co-founder of Audio Precision. “On the one hand, things can happen that are audible but don’t show up in a typical characterization procedure. Take, for instance, PC sound cards. Traditional measurements might say that a card performs splendidly but would never tell you that, when the PC is busy, the software will miss a deadline and cause a dropout that is very audible to listeners. On the other hand, I really want to believe that if something is audible, we should be able to measure it somehow. It takes measuring and listening.”
Manufacturing test
If characterization is such a complex question, manufacturing test in the SOC world promises to be a nightmare. “The key issue here is that you are trying to ensure the quality of a part, and you have a total of 5.5 seconds of test time,” explains Marcel Tromp, engineering fellow at LSI Logic. That total time budget is insufficient to complete some of the individual tests that are routine on the characterization bench, let alone thoroughly test an analog output. And one home-theater SOC might have nearly a dozen outputs. Add to this problem the realities of the manufacturing-test environment and the variety of customer-test requirements. “Some customers nearly ignore the ability of the chips to deliver fantastic audio quality; they just don’t care as long as the output is functional,” laments Wolfson’s Frith. “And then there are others, like the Japanese systems manufacturers and the automotive industry, who test everything,” adds Hayes. “We have some customers who just have a scope on the end of their production line and others who measure every incoming chip with an Audio Precision box.”
Just performing manufacturing tests is becoming a challenge (Figure 2). Frith says that testing analog from a 24-bit, 192k-sample source can take all of the dynamic range—and, more challenging, all of the noise margin—of which the best Teradyne mixed-signal testers are capable. And the paucity of high-dynamic-range signal cards can mean that engineers will test analog outputs sequentially, rather than in parallel groups. But even this problem may not be the most serious one. “At this level, there is a trend for the R&D department to specify the whole test environment,” Hofer says. “But that becomes difficult in today’s world of third-party testing houses where distance and culture are barriers. In China, I’ve seen facilities in which you couldn’t count on the third wire for the test systems’ actually being grounded anywhere.” This news is dreadful for an environment that is high in electrical noise under the best circumstances.
So test design becomes an art. Characterization engineers must work with test engineers to find a minimum number of tests—of which the real-world test equipment will be capable within the time budget—that will provide a high probability that the chip will sound as the customer expects. Only experience, a thorough knowledge of audio, and good luck can achieve that. “Defining a set of quantitative tests is becoming a skill set in its own right,” LSI’s Tromp says. “How do you bring nonquantitative ideas of good and bad into the engineering world? It’s much the same problem as in video, where the ultimate judge is the viewer. But, at least in video, if something doesn’t look right, you can stop the frame and examine it.”
There are some hints from the experts, though. “The majority of test problems come down to routing and accessibility,” Hofer declares. “Accessibility is key. For instance, if you have an on-chip DAC, you need to be able to get to both sides of it. Otherwise, you are just measuring the subsystem end to end with no idea of what’s going on inside.”
This issue makes design for accessibility a critical skill for high-end audio. Engineers need to bring test points out of the SOC. But the routing of these test signals is every bit as critical as the routing of the real analog output, or the data will be nearly useless. Crosstalk, an analog multiplexer with a high noise floor, and any number of things that seem unimportant can destroy the visibility into an analog node when you are making measurements with 110-dB dynamic range.
Ironically, a concept from digital self-test may be important here, as well: structural, as opposed to functional, test. Given that there simply won’t be enough time to fully test the function of a precision-audio output, the designer must understand the likely failure modes and design the chip to help inspect for them. “We are fairly fortunate in that our stuff is digital until it hits the second-order filter,” says TI’s Shipley of the company’s Class D amplifiers (Figure 3). “But even so, we have to relate what is happening in the analog signal back to the digital architecture, based on our understanding of the circuit. The tester looks at switching waveforms on the die, and you have to be able to know how that impacts the sound the customer will hear.”
In the end, manufacturing-test procedure becomes a cooperative effort between the SOC designer and the test system. Perform a few specific tests with the test system and find out how this chip will sound driving a pair of headphones or a Class D amplifier from Brand X. It’s no small challenge, but with listening rooms the ultimate arbiters of quality at the high end of the audio world, it’s a challenge that SOC design and test engineers must face.
CAN YOU EVEN MEASURE THAT
Unfortunately for audio engineers, many sorts of artifacts are clearly audible to a trained listener yet completely transparent to traditional audio-characterization procedures. A few examples might illustrate the subtlety with which a designer must arbitrate between the listening room and the characterization bench. One is a development at Wolfson. Bench tests showed that a part was excelling; it had wide dynamic range and low distortion. But listeners said that the sound stage was inaccurate. Further exploration showed that the cause of this problem was the digital-filter algorithms for FIR (finite-impulse-response) filtering. The most common algorithm does exactly the right thing in the frequency domain, in which engineers check the filter response. However, observing the impulse response in the time domain shows that the response centers on—rather than trails—the impulse. In other words, preringing occurs. This effect interferes with the human ear’s method for keeping track of the spatial relationships between the sounds from the speakers. A new FIR algorithm was in order.
In another example, the scrambling algorithm in high-precision DACs becomes stuck in cycles, causing random—and, hence, inaudible—spikes to group into audible repetitive sequences. Traditional characterization does not reveal the problem, but careful linearity testing over a long data sequence does.
You must even respect the implausible listening-room result. This area is particularly sensitive, because, according to many engineers, some listeners claim to hear things that are neither reproducible nor real. But it is imprudent to dismiss a listening result just because it sounds wacky.
Morten Lave, chief executive officer of TC Applied Technologies, offers the following example: A listening test compared the audio quality of a CD player, amplifier, and speaker combination using an analog-through-RCA-plug interconnection scheme and an S/PDIF (Sony/Philips-digital-interface)-through-optical-fiber scheme. Listeners reported that the analog connection produced better sound. The report on the test concluded that this finding was yet another proof of the superiority of good old analog over digital. Lave initially ignored the result.
But looking further, he found that the investigators had conducted a blind test in which the listeners didn’t know which system they were hearing on which passages, and the results were consistent. So, engineers dug further and found that measurable differences existed between the rise and fall times of the optical transducers in the S/PDIF optical link. This difference caused data-dependent jitter, which, on the other side of the DAC, had an audible component. “I believe in blind tests,” Lave says, “but not always in the explanations.”
AT A GLANCE
New digital-media standards are bringing high-fidelity audio to SOC (system-on-chip) platforms.
High fidelity presents a new kind of characterization problem to SOC designers: one that they can't always solve quantitatively.
Manufacturing test for these chips will be a gamble based on art and architecture.
As users' expectations for high fidelity increase, design for test will take on a new importance for audio functions.
FOR MORE INFORMATION
Audio Precision: www.ap.com
Dolby: www.dolby.com
LSI Logic: www.lsil.com
National Semiconductor: www.national.com
Nvidia: www.nvidia.com
TC Applied Technologies: www.tctechnologies.tc
Teradyne: www.teradyne.com
Texas Instruments: www.ti.com
THX: www.thx.com
Wolfson Microelectronics: www.wolfsonmicro.com
Illustrations:
Figure 1Figure 2Figure 3
