|Home | Audio Magazine | Stereo Review magazine | Good Sound | Troubleshooting|
ENDING THE REIGN OF ERROR
Error protection is one of the great opportunities of digital audio storage media such as the Compact Disc. It offers something that was never possible with analog media-the chance to correct errors and repair damage.
When you scratch an LP record, the grooves are irrevocably damaged, along with the information contained in them. Forever after, there will be a click or pop as the damaged part of the groove passes underneath the phono graph stylus. But when you scratch a CD, the nature of the data on the disc and the player's design offer you a second chance. The data on the disc includes a special error-correction code, and your player uses this to correct for damaged data. Thus, it strives to deliver the original information in tact, and it performs this error correction each time the disc is played.
Let's examine some fundamentals of error protection. To start off, consider these two messages:
1. Turn to page 345 in your hymnals.
2. Let's party! Let's party!
The first message is like the information in an LP groove, and the second is like the data in a CD pit track. Now, lay a finger vertically across the two messages, to represent a scratch. You'll observe that part of the first message is irrevocably gone, whereas the second message can be reconstructed because we have used redundant data to repeat it. In fact, redundancy is the essence of error protection. This is extra information, derived from the original data, which is not essential (in the absence of errors) to the message itself. In general, the greater the redundancy, the greater the error protection.
By comparing redundant messages, you can overcome the effect of the error. Note that we haven't prevented the error; we have simply ensured against its effect. Clever, eh? Of course, even the most clever of ideas has its limitations. As you lay more and more fingers over the message, reconstructing it becomes more and more difficult. Thus, the severity of the error is an important factor in our ability to make corrections. Also, if you lay your finger horizontally across the page, you might completely obscure the entire line, destroying both the message and its redundancy. There fore, the nature of the error plays a role too. In addition, we should note that error correction demands a price; in this case, including an error-correction code doubles the amount of storage space required.
On the negative side, error protection is not only an opportunity but an obligation. Digital audio on a Compact Disc might require storage of 15 billion bits, which amounts to about 100 mil lion bits per square centimeter. With such great data density, even the smallest speck of dust would wipe out a considerable number of bits. This points out the fact that, in reality, the type of error correction utilized has to be a great deal more sophisticated than our simple redundancy scheme.
We should also consider the vulnerability of digital data. Even one bad bit could wreak havoc. For example, if the digital word 0000000000000000 (representing silence) was misread as 1000000000000000 (representing a pretty loud level), a loud click would result. Highly unacceptable! Thus we have no choice; for satisfactory digital audio storage, error protection is essential. Data on a CD must be error-protected before the disc is made, and a CD player must be able to correct errors during playback.
But before a CD player can correct errors, it must detect them. While this might sound obvious, the problem can be considerable. If presented with a data word, could you tell whether or not it contained an error? For example, does 1100101000011110 contain an error? Unless you are psychic, there's no way to tell. I could transmit the message twice:
and close examination would reveal a difference between them. Obviously, both cannot be correct, but which message is right? I could transmit the message three times, with two identical data words and one slightly different, and you might have a good suspicion of which was correct. But would you be sure? What if I repeated the message three times and all three versions were different? Obviously, simple repetition is an in efficient way of going about error detection and correction; we need to de vise a better method. Redundancy is the best way, but we must optimize it to limit its inherent penalty of overhead.
This leads to the development of elaborate error-correction codes which make very efficient use of redundancy
by coding the information in certain ways. Error-correction codes use special algorithms to create redundant in formation, balancing the information between the twin tasks of detection and correction.
What is a code? A code is simply a way of representing information. Common examples include Morse code, ASCII (American Standard Code for In formation Interchange), and the ISBN (International Standard Book Number) code. An ISBN number is more than just a series of digits. For example, consider the ISBN number 0-672 22388-0. The first digit designates the country of origin; for example, 0 is for the U.S. and some other English speaking countries. The next three dig its comprise a publisher code, and the next six make up the book number code. The last digit is particularly interesting; it is a check digit with a value derived from the values of the previous digits. This allows us to verify the correctness of any ISBN by adding together the first nine digits, in modulo 11 arithmetic, and comparing them to the last digit.
In practice, the redundancy con tamed in digital audio protection codes acts very much like this last digit of the ISBN. Specifically, the redundancy of ten takes the form of a parity bit, an extra bit chosen according to a predefined rule, using the message as the basis. For example, in an odd-parity scheme, the parity bit is chosen so that the number of ones in the message, including the parity bit, is odd. A word such as 11010100 would have a parity bit of 1 appended. When the word is read from a storage medium, we can use the parity bit to perform a simple check on the word's validity. In this case, if the number of ones in the received message were even, we would know an error had occurred, though not which bit was wrong.
To increase their detection reliability, parity codes can be made more complex. In some codes, data is divided into blocks and parity bits are added to each block. Consider the example in Fig. 1. In addition to the 16 data values, nine extra parity values are created and appended. They are placed at the end of each row and column, and each represents the sum of that row or column. Note that a parity value is also given for the parity row and the parity column.
If an error occurs in any data (or parity) value, the error can be easily located and the correct value can be easily calculated, using the other data present. For example, suppose that the data block shown in Fig. 2 is received. As we calculate each parity value at the receiving end, and check it against each transmitted parity value, we observe a disagreement. In fact, there is a disagreement in both a row and a column parity value. The intersection of the row and column points to the error. Furthermore, we can now substitute the correct value, thanks to the transmitted parity values.
The data in the block is thus made substantially more reliable than it had been. Of course, instead of 16 values, 25 are now required in order to accommodate the error code.
Using parity, the system can- detect and correct errors, and supply good as-new information. However, as I've mentioned, the performance of the error-protection system depends on the nature of the error. What if a large error obliterated both the data and its parity? There would be nothing left from which to reconstruct the message. It is thus important to understand the nature of the errors and devise protection accordingly.
With interleaving (8). the effects of dropouts are distributed, allowing errors to be corrected or concealed.
Errors can occur in large groups, called burst errors, or in isolated instances, called random errors. For ex ample, a badly formed pit would cause a random error, whereas a greasy fingerprint would cause burst errors. The CD must guard against both kinds.
Interleaving is employed to protect against the all-too-likely occurrence of burst errors. Interleaving might be thought of as shuffling a deck of cards; data words are redistributed in the bit stream during recording so that consecutive words are never adjacent in the medium. An error occurring in the medium (such as a dust particle on the disc) might prevent the successful reading of a number of adjacent values. ues. However, upon de interleaving, the shuffled words are placed back in their original and rightful positions in the stream, and the errors are scattered in time. Thus isolated, they are much easier to correct. Figure 3 shows what happens to a data dropout without and with interleaving. In the latter case, a large defect in the medium is distributed following de-interleaving.
Because consecutive data bits are stored far apart, any errors will be scattered, becoming more like random errors-which are more easily corrected.
Figure 4 shows how interleaving and de-interleaving works. It appears complicated, but simply delaying the data with differing delay times during re cording, and delaying it again (in a complementary manner) upon play back, accomplishes the job. Cross-interleaving carries the idea one step farther; data is interleaved, then inter leaved again. This adds extra robustness, providing correctability for larger errors. The nice part about cross-interleaving is that while error protection is enhanced, the amount of redundancy is not increased.
While total error correction is theoretically possible (short of catastrophic failure of the system), it would be impractical to implement. With real-life digital audio systems, some errors are too massive for the error-correction scheme, and we must fall back on error concealment. Using error-detection results to pinpoint massive errors, interpolation methods utilize valid data surrounding the error to calculate a new value to substitute for the error.
Unlike correction, in which the new value is identical to the original, interpolation is only a best guess. But with a good interpolation scheme, an uncorrected error can be made virtually in audible.
In worst-case scenarios, where the error is so massive that even error concealment would fail, we choose to mute the audio signal. (After all, brief silence is preferable to a burst of distortion.) By swiftly attenuating the signal before and after the mute, even these catastrophic errors are often made inaudible to most listeners.
As we have seen, error protection consists of several kinds of processing tasks. Errors must be detected and then corrected (if possible), concealed, or muted. To accomplish those chores, the signal must be carefully encoded with detection and correction codes, and interleaved. It's a lot of work. Next time you're eating some thing greasy and listening to CDs, I hope you'll appreciate it.
(adapted from Audio magazine, May 1987)
= = = =