Synthesis snd/sp‎ > ‎

Audio processing

dropbox


"adding a Virtual Tape Loop"

source: Byte Publications, june 1978

by Tom O'Haver, Dept ofChemistry, University of Maryland, College Park MD 20742


  • There is a lot of talk about digital audio processing, but talk is not the same as practical action. As the prices of microprocessor systems and interface devices continue to drop, such applications are sure to become quite common, even among amateurs. This article describes a few of the possibilities of the use of a small low cost microprocessor system for djgital processing of audio signals. The effects described involve echo, reverb, fuzz, time delay, phase phlanging, mono-to-enhanced-stereo conversion, and frequency multiplication. These effects could also be quite useful for the experimentally inclined audio enthusiast or music group.

Hardware Requirements

  • To run the programs given here, you will need a 6502 (or equivalent) processor with from 1 K to 5 K bytes of programmable memory, an 8 bit input port connected to a fast 8 bit analog to digital converter (ADC), and a latched 8 bit output port connected to an 8 bit digital to analog converter (DAC). An additional output port and digital to analog converter are required for stereo applications. The basic hookup for a simple monaural system is shown in figure 1.The signal from the preamp is amplified, low pass filtered, converted to digital by the analog to digital converter, processed in the microcomputer, converted back into. analog by the digital to analog converter, and then filtered some more before going to the power amp. The success of such a system in audio processing depends upon its ability to operate at ultrasonic speeds; that is, the rate at which the audio signal is digitized, processed, and output must (or should) be as far above the upper limit of the audio spectrum as possible. Thus the speed of each of these steps is critical. We'll consider each step individually.

    Next to the microcomputer itself, the analog to digital converter is really the most critical component. It must be a fast one; a conversion time of 50 ,us or less is necessary to allow sufficiently high sampling rates.
    I have been using a Datel Model E8HB1, an 8 bit successive approximation analog to digital converter with a 4 us conversion time (available from Datel Systems Inc, 1020G Turnpike St, Building S, Canton MA 02021, for $85 in unit quantities). I recommend this unit. Its conversion time is probably faster than you will need, but at least you won't have to buy a new one when microsystems get faster (as they certainly will). In addition to speed, the Datel analog to digital converter has two other features you should look for in a converter. First, it is bipolar, which means it is capable of accepting both the positive and the negative excursions of the audio waveform. Otherwise, you would have to add some offset to the incoming signal. (The programs in this article assume offset binary coding.) Second, it clips on overload rather than wrapping around. That is, if the input audio signal exceeds the dynamic range of the analog to digital converter, the digital output simply stops at full scale rather than wrapping around or folding back to zero. The reason this is a useful feature is the fact that an audio signal contains many peaks and transients which greatly exceed the average signal amplitude. With only an 8 bit system, there is really no way to keep these transients from exceeding the range of the converter, at least occasionally; if you try to prevent it by adjusting the average amplitude to a very low level, you'll get too much quantization noise. Clipping the peaks may offend the audio purist, but I'll guarantee you that it sounds a lot better than wrapping around. Of course, a better solution would be to use 12 bit converters and a 12 or 16 bit (or faster 8 bit) computer. Sufficiently fast 12 bit analog to digital converters are available for about $150, and 12 bit digital to analog converters are typically about $30. But without a 12 or 16 bit processor, all processing would have to be done in double precision, which might slow things down too much (unless you have a 4 MHz 6502, which I do not). Anyway, an 8 bit converter which clips is good enough for the time being.

    Selection of a digital to analog converter is much easier, since several fast, low cost 8 bit units are available. The digital to analog converter needn't be bipolar, since a DC blocking capacitor can be added easily. The Hybrid Systems 371-8 at $10 is a good choice, as is the Motorola MC 1408L8 at about $5. I've used both successfully. The Hybrid Systems unit is more convenient because it has a built-in reference supply, while you will have to supply an external (2 V) reference for the Motorola unit. This must be very well filtered but not necessarily well regulated for audio applications. (An advantage of the Motorola units is that they can be used as multiplying digital to analog converters. If you drive the reference input of one converter from the output of another converter, then the output of the first con- verter will be the product of the digital inputs to the two converters. This allows you to obtain automatic level control, compression, expansion, fading, and ampli- tude modulation effects without relying on much slower software multiplication routines and without getting into trouble with quantization noise.)

    As for the processor itself, almost any 6502 system should do with the examples I've included in this article: KIM-1, Jolt, Ebka, 0SI, PAIA, Apple-II, PET 2001, etc. I've used both the Ebka and the 0SI systems with good results. 0SI has a particularly convenient analog IO board (Model 430) which can be populated with two MC1408L8 8 bit digital to analog converters, an 8 bit analog to digital converter, and their associated latches and address decoding logic. The 0SI Model 430 analog to digital converter circuit is of the synchronous tracking (up- down counter) type. Be warned, however, that this analog to digital converter wraps around on overrange. It also requires some individual tweaking of component values to get it to work. If you want to use the OSI 430 board, I strongly recommend that you replace their analog to digital circuit with a better one, such as the Datel E8HB1. Other than that, the OSI board is just fine.

    One very important concept which you must understand is the relation between sample rate, aliasing, and low pass filter ing. If you don't understand these terms and their significance, then before you go on you should read the article by Hal Chamberlin on page 62 of the September 1977 issue of BYTE. For the programs presented in this article, the sampling rate will fall between 20 and 40 kHz with a 1 MHz processor clock frequency, assuming you are using a sufficiently fast analog to digital converter (less than 50 us sampling time). To control aliasing, you have to roll off the high frequency response of the input signal to the analog to digital converter at a frequency no higher than about 1/4 of the sampling frequency, ie: about 5 kHz for a 20 HZ sampling rate. This may not sound muc like "hi-fi," but actually it sounds better than you might think. For better highs, you need faster processing and maybe a faster input converter. The 6502 is pretty good in this respect; it's available in versions at Ieast up to 4 MHz. This would give you a sampling rate of 80 to 160 kHz for the programs given here and would extend the highs to the 20 to 40 kHz range. Now, that's a high fidelity computer!

    The sampling rate therefore determines the frequency at which the response of the system m'ust be rolled off (by means of appropriate low pass filters) in order to reduce aliasing to a tolerable level. In the simple circuit of figure 1, the only roll off is that provided by the capacitors in the feedback loops of the two op amps. Al- though this circuit is satisfactory for experimental purposes, the cutoff rate of the high frequency rolloff is not sharp enough for first class results. If you're really serious, you'll want more sophisticated, sharp cutoff filters. Hal Chamberlin gives the circuit of an excellent filter in his article in September 1977 BYTE, mentioned previously. The unpopulated printed circuit board, as well as an assembled and tested unit, is available from Hal. The cutoff frequency of this filter is 3 kHz, probably too low if you have a reasonably fast processor, so you might want to modify it or roll your own based on the designs in Don lancaster's Active Filter Cookbook or other reference sources. Only one sharp cut filter is needed, between the preamplifier and the input analog to digital converter, to reduce aliasing. The filter on the output, between the digital to analog converter and the power

    Figure 1.. The system design of an audio processing test bed requires two simple peripheral devices and a computer. The input device is an analog to digital converter (ADC in this diagram) preceded by a filter. The output device is a digital to analog converter (DAC in this diagram) driving another filter. Source material from (for example) a broadcast program is input through the ADC, processed in real time by the program in the computer, then output in real time to the DAC where it (for example) goes to your audio power amplifier and speaker system. The program in the computer can be as simple as an unprocessed transfer from input to output, or as complex a transfer function as the constraints of real time will allow, given the speed of the computer.



    amplifier, needn't .have as sharp a cutoff, but it should have the same cutoff frequency as tht input filter. For the simple filters in figure 1, the cutoff frequency is equal to 1/2nRC, where R and C are the values of the feedback resistor and capacitor, respectively.
    One last thing to consider about the hardware is the level (amplitude) of the audio signal. In order to avoid excessive quantization noise, the input signal must be amplified enough to utilize the whole dynamic range of the input analog to digital converter. In figure 1, the first op amp provides gain in addition to filtering. The gain of this amplifier, which .is equal to R2/R1, will have to be adjusted for your particular system. Given choices of R and C for filter cutoff, R1 can be chosen given a desired gain level. For example, if your preamp provides a maximum output signal of 0.2 or 1 V peak- to-peak, and your input converter has an irlput voltage range of abt 5 V (10 V peak-to- peak); then a gain of 10 V/1 V = 10 is appropriate. Also, the maximum output signal of the digital to analog converter must not be allowed to overload the power amplifier. This will dictate the selection of the feedback resistor R0 of the output op amp in figure 1; the output voltage is directly proportional to the value of th is resistor.

Software Considerations

  • So what about the software? First, let's see how to get data in from the input converter and out to the output converter without any processing at all. If your analog to digital conversion device (which I reference symbolically as CONV) is connected to an input port whose address is F800, then to load one sample of the audio signal into the accumulator (A) register of a 6502 requires one instruction, thus: LDA CONV
    This is all you need if you're using a tracking converter such as that on the OSI board, but if you're using a strobed converter, you'll have to give the converter a strobe pulse first, allow it time to convert, then load the A register. The fastest way to do this is to assign the input conversion strobe to an unused address, decode that address, and use the address select line (address "strobe," as it is sometimes called) as the pulse which strobes the converter. I have used the latter approach for my strobed analog to digital converter and have (arbitrarily) assigned address EC00 (which I call STROBE symbolically) to the address strobe. With this arrangement the converter is strobed by any instruction which references that address; for example an STA as shown here:
    STA STROBE
    -
    - several instructions executed while conversion occurs
    -
    LDA CONY

    The above routine strobes the converter and then loads the data into the A register. The dashes represent intervening instructions which take up enough time to allow the analog to digital converter to complete its conversion. This will always be a useful code, rather than just no operation instructions (NOPs) or a wait loop. Conversion times of commercial analog to digital converters vary allover the place. As I mentioned before, for audio processing you'll need a fast one which converts in a time of 50 uS better. Just make sure there are enough instructions between referencing STROBE and the loading from CONY to give the input converter time to convert.
    To output one sample to the digital to analog converter (called DAC symbolically) is quite simple. For example, if the converter is connected to an output port whose ad- dress is F900, then all I have to do is store the sample: 8D 00 F9 STA DAC

    To test the proper operation of the input and output converters we can write a "straight wire" program which simply transfers the data from the input to the output without change. Listing 1 shows 6502 position independent code for such a program.
    Note that the input conversion strobe instruction is placed right after the load CONY instruction. This may seem backwards but it gives the analog to digital converter a total of seven machine cycles (an STA and a JMP) to convert before it will load into A. On a 1 MHz machine, this means the conversion time could be as long as 7 uS. If your converter is slower than this, put some NOP instructions or a wait loop right before the CLC instruction. The other programs in this article execute much more code between strobing and loading the input analog to digital converter and will usually allow you
    , to get by with no additional instructions ntended specifically to slow down execution.


    The program of listing 1 is good for testing out the hookup to your audio system. The sound quality of music played "through your computer" this way may be better than you would expect, considering that the audio waveforms are being sliced up into discrete samples, converted into binary numbers, and then converted back into an analog audio waveform!
    So what kind of audio processing can you do? I'll resist the temptation to say that the applications are limited only by your imag- ination. They are not. They are limited by your programming skill, your processor speed, and your system's programmable memory capacity. You can never have too much of these. I'll not claim to have even scratched the surface of potential applications in this article. I'll just tell you about a few things I've done, mostly because they were easy to program. If you don't come up with better ideas than these I'll be disappointed.

Waveform Modification



  • A very easy class of audio processing functions are those which are intended to distort the audio waveform. Believe it or not, distortion is actually considered desir- able by musicians in some cases for obtaining special effects (such as "fuzz") with electric guitars and other electronic instruments. The computer can perform a rather elegant general purpose distortion function by utilizing a stored transfer function as illustrated in the program of listing 2.


    To use this algorithm, you must set up a table in memory (on page 03 in this example) which serves as the transfer function. Each sample of the input waveform obtained from the input converter is used as an index to look up a corresponding byte in the table, which is then used as the output value. In this example the table is just 256 bytes long and is indexed by the 6502 processor's X register. Depending upon what we store in the table we can get any kind of distortion effect we want. A trivial case would be to use a straight line function, ie: put 00 in 0300, 01 in 0301, 02 in 0302. . .and FF in 03FF as shown in figure 2a. This would yield no effect at all; the output would be identical to the input as was the case with the prqgram of listing 1. But if we use anything other than a straight line, we'll get distortion. Several possibilities are shown in figure 2. Figure 2b would give a square wave output while 2c would yield a somewhat less strongly distorted output. With the function shown in figure 2d, we would get a frequency doubling effect; that is, if the input were a sine wave of one frequency, the output would be an approximate sine wave of twice that frequency (one octave higher). With figure 2e, we'd get an output two octaves higher. This can be extended even further with the appropriate transfer function (eg: figure 2f). The effect on the sound of an electric guitar is quite remarkable, particularly as the frequency multiplication factor is a function of the amplitude of the input signal and changeat the input decays. Quite apart from its potential uses in music recording or performance, the above technique is a neat way to teach (or learn) about the effect of transfer characteristic nonlinearity on audio distortion. Just put in the characteristic under consideration and listen to the effect it has on the audio.

Time Delay, Phase Shift and Reverb Effects

  • If we store digitized audio in an array in programmable memory, and read it out to the digital to analog converter at a later time, we have a time delay effect which can be used for phase shift and reverb. The maximum time delay you can achieve depends on the sampling rate and the amount of available memory; but even with only 256 bytes you can get some pretty good phase shift and phase "phlanging" effects. Witl 4 K bytes you can get a good reverb.

    The essential programming technique behind all of these effects is quite simple: output a byte from the data buffer to the digital to analog converter; input a new sam- ple from the analog to digital converter and put it in the same location in the data buffer as the byte just output; increment the pointer modulo the length of the buffer and repeat. (Thus, when you get to the end of the data buffer you reset the pointer to the beginning and continue.) By scaling and adding the new data from the input converter to the old data from the data buffer, we can generate a range of effects depending on the length of the time delay.



    The routine of listing 3 adds the audio signal to a slightly delayed version of itself and outputs the scaled sum to the digital to analog converter. In this routine, page 03
    serves as the data buffer and x as the pointer. The pointer is initialized to DELAY, decremented until it gets to zero, and then reset to DELAY. This results in a sort of circular data buffer which acts as a first in last out shift register. The new and old (delayed) data are added and sent to the output converter. (Note that to prevent overflow, the data are divided by two before adding.) The time delay, determined by DELAY, can be adjusted from 1 to 225 samples. Such short delays do not result in a perceptible echo. The effect is rather that of a "comb filter" with multiple peaks and dips distributed throughout the audio spectrum. This is due to the fact that there will be a cancellation at every frequency whose period is an integral multiple of twice the time delay and a reinforcement at every frequency whose period is
    an integral multiple of the time delay. This rearranges the amplitude and phase relationships of the harmonics of music and speech and has a quite noticeable effect on the sound, variously described as a "resonant" or "twangy" effect. (If you have a hum problem in your audio setup, you might try to find the value of DELAY which puts a dip right at the hum frequency.)

    The above idea can be extended and the effect made much stronger by causing DELAY to change continuously in real time. This would cause the peaks and dips to sweep through the audio spectrum. This effect is called "phase phlangjng" by some people. An easy way to do it (not necessarily the best way, however) is shown in listing 4. This is the same as the previous program except that the DEC DELAY instruction has been added to reset the buffer pointer to a different value each cycle through the buffer. The effect of this routine on voice and music is quite dramatic. With speech and solo singing it gives a kind of voice doubling effect, as if two people were speaking or singing in synchroniza- tion. It makes a 6 string guitar sound re- miniscent of a 12 string guitar. A concert piano comes out distinctly like a questionably tuned honky tonk piano. The effect on organ music is unreal and unpleasant. If you play the guitar and sing, or think you do, try processing a tape recording of yourself this way. It will sound better, or at least different (which in my case is the same thing).
    If you have two output ports and two digital to analog converters, you can generate two channels of audio output. For example, you can convert a monaural source to "pseudostereo" with a further modification of the program in listing 3 (see listing 5).

    In this example the additional DAC is connected to an output port whose address is EF00. Instead of being added together, the direct and delayed signals are simply sent to the two different channels. The result is a sort of stereo phase phlanging effect which sounds much like a "rechanneled for stereo" disk recording. Try this through stereo headphones. So now you can have a stereo electric guitar. Would anyone like to extend it to quadraphonic?
    If you have at least 4 K bytes of memory available in your system for your buffer, then you can obtain echo and reverberation effects quite readily. The idea is basically the same as the phase shifter routines just discussed, except that a much larger data buffer is used. Here we can use the indirect form of the LDA and STA instruction, and we maintain a 16 bit pointer in page zero (unlike the 6800, the 6502 has only 8 bit index registers). The routine of listing 6 yields a reverberation time which is adjustable up to about 0.5 seconds. The data buffer is assumed to be the 4 K byte block from addresses 2000 to 2FFF. On each cycle through the buffer, the old data is divided by two, added to the new data out- put, and returned to the buffer. Thus, the old signals (ie: the echo) die off by a factor of two each time they are heard. You can hear about five or six echos before they drop below audibility.

    If START is set to 20, using the whole 4 K buffer, the effect is something like that of a large hall or perhaps an old railroad terminal. The difference is that the com- puter produces a clear, clean echo at very precisely timed intervals and with a precisely controlled decay rate. Compare this with either a natural reverberation situation or a mechanical unit: the result is a more me- chanical sound, much like a tape loop reverb device, without the false resonances of a spring type device. The advantage over a tape loop device is, of course, that it will never wear out or get out of alignment.



    Several useful modifications of this program can be made. For example, you could utilize a second digital to analog output and a stereophonic sound system to achieve spacial separation between the direct and "reflected" sound. You could then apply some filtering to the reflected sound channel to simulate selective absorption by the room furnishings. You could also improve the real- ism of this effect by writing the routine to provide more than one delay time, for example by maintaining two or more buffer pointers which would allow the incoming data to be added to several points in the data buffer. You'll need a fast processor to keep the sampling frequency up, however. Finally, by simply dropping the LSR and ADC instructions in the program of listing 6, you can get a simple time delay effect; say a word and it is repeated immediately. Great for language study; listen to and critique your pronunciation without wearing out your tape recorder. Or if you have lots of memory (at least 32 K), you can get delays long enough to allow you to sing a round with yourself! I won't com- ment on the frightening social significance of this.




posted : 20 oktober 2002



Back To General Info   Back to Richard Davies Home Page
Comments