Creating Virtual Voices with Amiga Audio

by Dan Baker



Every Amiga model comes standard with 4-channel, 8-bit stereo audio
hardware.  This hardware provides every application with the capability
of producing 4-part, stereo sound.  Some applications however may want
to exceed the 4-channel limit.  For games and other applications that
use sound effects extensively, it may be desirable to trigger more than
4 sounds simultaneously.  This article demonstrates two techniques you
can use to implement virtual voices on the Amiga's audio hardware
effectively doubling the number of voices available to 8.


Audio Hardware Limits

Using DMA, the audio hardware can fetch about 2 bytes per scan line for
each channel without processor intervention.  This means that each
audio channel can play back sampled data at up to 28,867 bytes per
second.  It turns out that this 28,867 byte per second limit far
exceeds the requirements of most audio sample files.

The typical 8SVX file contains data sampled at a rate of about 10,000
bytes per second.  In that case, the audio hardware only uses about 1/3
of the available audio bandwidth.  The audio hardware is capable of
fetching much more sound data without affecting system performance.
The question then arises, can this extra horsepower be harnessed in
some way?  The answer is yes.



Audio Interleaving

An application can use any extra audio bandwidth to interleave the
bytes from two separate sample files.  This technique allows an
application to play both samples simultaneously on a single audio
channel.

For instance, if you have two files sampled at 10,000 bytes per second,
you could set the playback speed to 20,000 bytes per second and
alternate playing bytes from each sample.  The effective speed of each
sample is preserved since each sample is output only half the time.


 Figure 1 - Interleaved audio data 


As shown in the diagram above, an application interleaves the data
bytes from two sound files into a memory buffer.  The application then
plays the back the interleaved sample at twice the rate.  Note that for
this method to work, the speeds of the two samples should be the same,
or at least a close match.  Another restriction is that the sample
speed must be no more than half the maximum speed available on the
Amiga.  Since the playback speed has to be doubled for interleaving to
work and since the speed limit of the Amiga's built in audio hardware
is 28,867 bytes per second, the speed limit of each interleaved sample
is 14,433 bytes per second.

The biggest advantage of audio interleaving is that it exploits unused
audio bandwidth available on the Amiga hardware.  Also, interleaving
does not effect the values of the sampled sound data.  The individual
sound samples remain in their original, noninterleaved form. The
disadvantage is interleaving certain wave forms can combine into an
unexpected and undesirable wave form.  Consider the following two wave
forms:


 Figure 2 - Wave Forms 


When the Amiga interleaves these two wave forms, it has to alternate
between wave form A and wave form B.  Because the Amiga is constantly
oscillating between two wave forms, it produces a completely different
wave form:


 Figure 3 - Wave Form 


In practice, this usually isn't a problem if you are working with
sampled sound.  Typically, sampled sounds are not pure periodic wave
forms (like the wave forms above).


Audio Averaging

It is not always possible to interleave two samples.  For example, the
frequency of a sample may be more than 14,433 bytes per second, so
doubling it would exceed the 28,867 bytes per second limit.  In that
case, there is another trick you can do to the audio data to combine
two samples on one channel.

If the values from each sample file are added and the resulting value
is divided by two, a new data stream can be created which contains the
audio information from both files.


 Figure 4 - Averaged output data 


Note that in the example above, data values are represented as signed
bytes in the range -128 to +127 using twos complement format.  The
resulting average values are correct as shown.  Even though the
ioa_Data field of the IOAudio structure used for all audio.device
requests is shown as (UBYTE *) in the include file <devices/audio.h>,
do not be misled.  The data values are signed bytes.

Audio averaging introduces some noise into the resulting combined
signal.  Dividing the summed values by two effectively reduces the
dynamic range of the component samples by one bit (from eight to seven
bits).  Also, when the values are divided by two, any fractional amount
is truncated, hence, some information is lost.

Despite these drawbacks, the results of audio averaging on the Amiga's
eight bit audio hardware are comparable to the interleaving technique
described above.  In fact, the two techniques are virtually
indistinguishable on the current generation of Amigas (although this
may not always be true--see the ``Audio Experiments'' section later in
this article).


Virtual Voices

When two samples are combined on a single channel using interleaving or
averaging, both samples are clearly audible but, subjectively, it
sounds as if one bit of volume control has been lost on each sample.
Because of this, it is not wise to carry these virtual voice techniques
to an extreme.  Combining samples without limit will result in a badly
degraded composite in which the component signals are no longer clearly
audible. It is however quite feasible to double the number of available
voices to 8 using interleaving or averaging techniques.  The loss of
fidelity with 8 virtual voices is quite tolerable.  The code listed
below shows how this can be implemented.


Using the Interplay Program

The program, named interplay.c, allows the playback of standard IFF
8SVX files in three different ways:

1.  Normal playback of a single file on one channel
2.  Interleaved playback of two files on one channel
3.  Averaged playback of two files on one channel

For normal playback of a single 8SVX file, enter the following command
at the Shell prompt:

    1>interplay sample.8svx

This feature allows you to find out how a sample sounds alone as
compared with its interleaved or averaged counterpart.

For playback of two 8SVX files on a single channel using interleaving,
enter the following command at the Shell prompt:

    1>interplay voice.8svx music.8svx

The program reads the two files, figures out which has the faster
sampling rate, and sets the audio device to twice that value.  If the
calculated rate exceeds the maximum of 28,867 bytes per second, then
interplay sets the speed to the maximum.  The program then interleaves
the data from the two files so that the bytes played by the audio
channel alternate from one file to the other.  If the data from one
file runs out before the other, any remaining data bytes are
interleaved with zero.

For playback of two 8SVX files on a single channel using averaging,
enter the following command at the Shell prompt:

    1>interplay voice.8svx music.8svx SUM

The ``SUM'' keyword enables averaging instead of interleaving.  In this
case, the speed is set to whichever file uses the faster playback rate.
One byte is taken from each file, the two bytes are added and then
divided by two.  The resulting average value is played back.  If the
data from one file runs out before the other, any remaining data bytes
are averaged with zero.

Interplay can play samples of any size.  If a sample is too long, you
can terminate playback by pressing Ctrl-C.


How Interplay Works

Interplay uses a double-buffered approach for the playback of samples
of arbitrary length.  While one data buffer is playing, the other data
buffer is being prepared using either the averaging or interleaving
technique descried above.

Most of the code in the main loop within main() is concerned with
switching between one of the two playback buffers and their
corresponding I/O request blocks and message ports.  The averaging or
interleaving of bytes actually takes place in the FillAudio()
subroutine, not in main().

The reading and parsing of the 8SVX file are handled by the Parse8svx()
subroutine which takes as a parameter an InterPlay structure.  The
InterPlay struture holds all the state information that the program
needs to manage playback of the sampled data.  Thus there will be one
filled-in InterPlay structure for each file to be played back.  If the
user requests a combined playback, the two InterPlay structures are
linked together via the InterPlay.next_iplay field.  Otherwise this
field is set to NULL.

Housekeeping for the audio.device channels used is handled by the
SiezeChannel() and ReleaseChannel() subroutines.


Audio Experiments

Using the interplay.c program listed below, we found that there was
very little difference in the audio quality between the two methods of
combining samples.  We also found that for best results, the dynamic
ranges within the samples themselves had to be closely matched or the
result would be one sample drowning out the other.  Of course, it also
helps if the speeds are a close match.  If they aren't then one or the
other of the samples will be too slow or too fast.

Although both of these methods work comparitively well on the Amiga's
eight-bit audio hardware, doing the same tricks with 16-bit samples on
16-bit hardware would yield a different result.  As mentioned earlier,
the drawback of the averaging method is it loses a bit from the dynamic
range of the sound samples.  Compared to eight-bit sound, this loss is
much less significant when working with 16-bit sound.  As the dynamic
range increases, the impact of losing a single bit from the dynamic
range decreases.  On the other hand, the drawback to the interleaving
method is that it has to oscillate between two samples, which produces
a waveform equal to the sample playback rate.  The waveform is
independent of the dynamic range, so it remains constant as the dynamic
range increases.  The result is the avaeraging method will produce
superior results on systems with greater dynamic range.

Perhaps the best thing about the audio techniques demonstrated by
interplay.c is that they are not limited to the Amiga architecture.  In
fact, you can use interleaving and averaging with any system that
supports the variable speed playback of digitally sampled audio.  These
methods will work not only in the current generation of Amigas but in
any future system that support digital audio, although as the dynamic
range increases, the additive method will provide superior sound
quality.