** from an article by Philippe Rivaillon (privaillonfree.fr) in "La Voix de Paula" - Amiga News Tech issues 29-30 jan/feb 1992 (translated from french with some little summaries and omissions but not very important) :

Once upon a time,

"[...]It's Paula and incidentally Agnus for the transfers DMA which manage the sound. But its capacities arent only a simple spit of samples. Indeed, this diva is able to modulate the amplitude and the frequency of sound of one or more channels in order to produce the desired effects.[...]

A sound is a vibration, defined by two criteria: its amplitude and its frequency. The latter will determine whether a sound is low or high-pitched (the higher the frequency is, more the sound is high pitched).
The average amplitude defines the power of a sound (its volume), while the variations of this value define the sound itself. The data table transmitted to Paula will be thus a succession of instantaneous amplitudes. this table is cut out in bytes, for an amplitude from +127 to -128. Paula will of course have needs other information on the sound before playing it: its address, its length, its period (corresponding to the frequency of the signal) and its volume.

registers used :
*=# of channel (0-3)
+=channels 0=$A, 1=$B, 2=$C, 3=$D

AUD*LCH $dff0+0 Address of data in Chip
AUD*LCL $dff0+2

AUD*LEN $dff0+4 lenght of sample in words (16 bits) . max=128 kb

AUD*VOL $dff0+8 volume. 0-63. in words (16 bits)
only the 6 lower bits used.

AUD*PER $dff0+6 periode. speed of sample play. in cycle (16 bits)
frequency in bytes : Periode=1/(Freq)/279,365.10^-9 (for a 68000)

moreover than the audio circuit has limits because of its synchronization with the screen: knowing that it load a data word per rasterline, we obtains a maximum theoretical frequency of: 2 bytes*262.5 lines/screen * 59.94 screen/sec) = 31469 bytes/s
but the real max freq is 28867 bytes/s = periode 124 cycles.

If you put a value lower than 124 in this register, Agnus will not inevitably have time to charge the audio data and it's the old one which will be used, then a deformation of the sound.

First: simple transmission in automatic mode with Paula

summary:
- setup the registers above for each channel.
- to start the sound, enable DMA channels. (bit 9 and 0-3)
- Paula loop the sample automatically : after having read the number of words indicated in AUD*LEN, Paula will replace itself automatically at the beginning of the data.
- to stop the sound, disable DMA channels. if you give these channels DMA on the way, Paula will start again the sound from the beginning, with an exception : if the interval of time passed between the cut and the re-starting is lower than 2 times the period, then the reading will continue where Paula had left it.[...]

a quick example 1 :
start:
lea custom,a6
move.l #data,aud0lch(a6) ; $dff0a0
move.w #1,aud0len(a6) ; len in word - $dff0a4
move.w #5000,aud0per(a6) ; periode - $dff0a6
move.w #50,aud0vol(a6) ; volume - $dff0a8
move.w #D_setclr|D_Master|D_aud0,dmacon(a6) ; start dma channel 0
wait: btst #6,$bfe001
bne.s wait
move.w #D_aud0,dmacon(a6) ; stop
rts
data: dc.b 120,-120 ; square signal (in chipram)

my Example : 4 channels - 4 diff. period
lea custom+aud0lch,a0
move.w #3000,d0 ;period
lea Sample,a1
moveq #3,d1
loop:
move.l a1,(a0)+ ad
move.w #13,(a0)+ len
move.w d0,(a0)+ per
move.w #64,(a0) vol
addq.w #4,d0 ; incr period
addq.l #8,a0 ; next channel
dbf d1,loop
move.w #D_setclr|D_Master|D_aud,custom+dmacon
mouse: btst #6,$bfe001
bne mouse
move.w #D_aud,custom+dmacon
rts

Sample
dc.w $808a,$949e,$a8d2,$bcd0
dc.w $daea,$eef8,$d20c,$1c20
dc.w $2a34,$3e48,525c,$6670,$7a40 ; in chip"

to be continued...

Second: effects.

There are two kinds of possible modulations: amplitude modulation (volume) and frequency modulation (period).
The first could be used to make effects of vibrato or tremolo type, imitation of distance of a sound...
The second allows simulating chords.
Moreover, it's possible to combine the two modulations.
one problem: to modulate a voice, it's needed to use two consecutive channels.
the first (low) is used for the envelope and the second (high) is used for the basic sound.
the registers are initialized normally (except volume of the modulation channel, not played).
for the modulation channel, the data table is not made up of bytes but of words either a volume (0-64) or a period.

If you want to make an amplitude and frequency modulation at the same time, the table must alternatively contain a volume and a period, in this order. In this last case, there will be two readings instead of one so that the modulations are done simultaneously and not one after the other. Not need to touch at the period to compensate.

Then it's necessary to link the two channels for the modulation with register ADKCON (!! it is also used for the disk drives !!).

ADKCON/R: $dff09e/10
15 set/clr
07 ATPER3 if 1, cut voice 3
06 ATPER2 if 1, voice 2 module voice 3
05 ATPER1 if 1, voice 1 module voice 2
04 ATPER0 if 1, voide 0 module voice 1
03 ATVOL3 idem
02 ATVOL2 ...
01 ATVOL1 ...
00 ATVOL0 ...

bits 7 and 3 are same effect : stop the voice 3, the higher and so cant modulate an other. But even if no sound, datas continues to be loaded by Agnus. these 2 bits are useless. we can make use of it to create a break up of the audio in voice 3, but there are other less constraining methods to do that (amplitude modulation with a table made of only 2 values: 0 and 64).
After that, enable DMA channels : the basic sound will be played on the highest voice of both. It will use as volume or frequency (or both) the words loaded at each period with the channel of modulation. These words are loaded at the address of volume or frequency of the modulated channel. if you stop the modulation, dont forget to reset the volume or the frequency of the modified channels and disable DMA of the modulated channels.
It's possible to use only one channel, by modifying the volume or the frequency "in one's hand", but no syncro and cpu fully used!

example2:
lea custom,a6
; freq modulation
move.l #datamod,aud0lch(a6) ; datamod2 for amplitude modul. / datamod3 for both
move.w #8,aud0len(a6) ; 16 for both mod
move.w #40000,aud0per(a6) ; periode
; basic sound
move.l #data,aud1lch(a6)
move.w #1,aud1len(a6)
move.w #5000,aud1per(a6)
move.w #50,aud1vol(a6)

move.w #$8010,adkcon(a6) ; adkcon = 0 + 1 / 8001 for amplitude / 8011 for both

move.w #$8203,dmacon(a6) ; start
mouse: btst #6,$bfe001
bne.s mouse
move.w #$11,adkcon(a6) ; end of modulation
move.w #3,dmacon(a6)
rts
data: dc.b 120,-120 ; square signal
datamod:
dc.w 1000,2000,3000,4000,5000,6000,7000,8000 ; freq modulation
datamod2:
dc.w 2,8,10,20,30,40,50,60 ; amplitude modulation
datamod3:
dc.w 2,1000,8,2000,10,3000,20,4000 ; both
dc.w 30,5000,40,6000,50,7000,60,8000"

Third: interruption mode

[...]Amiga has 4 audio circuits, one for each voice. Agnus uses 4 DMA channels to transfer datas.[... ]

After the authorization of DMA transfers, Paula transfer address and length of the sound in its own registers. Afterwards, it generates an interruption of level 4 to inform the 68000 that it finished the operation.
The processor can then change the contents of the address and volume registers without that influencing the audio in progress. During this time Agnus load a word both periodes in register AUD*DAT $dff0+A

A cut of DMA channel is without effect if it lasts less time than twice the period of the played sound, because the state of DMA channel is checked only before each transfer of data, time between two transfers is the double of the period of the sound since the data are loaded word by word.

When the end of the sample reached, Paula reload audio registers and causes an interruption, until DMA stopped.
first problem: for example, we want to launch at the same time the same sound on the 4 channels. if the channels are not synchronized between them, that will produce a cacophony. But that will occur however only in one precise case: if the computer played already a sound on one or several of the voices. In this case, Paula will take account of the new sound only after having finished that which it was playing, from where a shift compared to the sound which began immediately.

Before any work with Paula, make sure that previous DMA transfers are finished. For that, it's necessary to wait time that the longest period lasts twice, 65535 cycles, or 20 ms with a CIA timer, more compatible with all amiga .

second problem: how to join 2 samples too long to play in one time ?
launch the first piece normally. After having loaded the contents of the registers, Paula causes an interruption of level 4. It's there that the interruption should be grabed, re-init address and length of the second piece without modifying DMACON.
During the second reading, these are the lately initialized registers that Paula will load, and it will be thus the new piece which will be played. If you have more than two pieces, the system is the same.

[skip note about interruptions in general]

example:
; save values
;...
; init
lea custom,a6
move.w #$7fff,intena(a6)
move.l #intlev4,$70.w
move.w #$c080,intena(a6) ; channel 0
; first sample
move.w #500,aud0per(a6)
move.w #50,aud0vol(a6)
move.l #data1,aud0lch(a6)
move.w #50000,aud0len(a6)
move.w #$8201,dmacon(a6)
mouse: btst #6,$bfe001
bne.s mouse
move.w #1,dmacon(a6)
; restore
;...
rts
intlev4: movem.l d0/a0/a6,-(sp)
lea custom,a6
move.w intreqr(a6),d0
btst #7,d0
beq .no
move.l pt,a0
move.l (a0)+,aud0lch(a6)
move.w (a0)+,aud0len(a6)
cmp.l #endtable,a0
bne .skip
lea table,a0
.skip:
move.l a0,pt
move.w #$80,intreq(a6)
.no:
movem.l (sp)+,d0/a0/a6
rte

data1: dcb.b 384000,0
table:
dc.l data1
dc.w 50000
dc.l data1+100000
dc.w 50000
dc.l data1+200000
dc.w 50000
dc.l data1+300000
dc.w 42000
fin:

pt: dc.l table+6

note: if the following piece doesnt have the same frequency or if you want to change his volume, you will have to wait the beginning of this piece (loading of the registers) to carry out the changes because volume and frequency arent safeguarded by Paula, who works directly on.

Fourth: manual mode

no more need to use audlch, audlen, nor to authorize the DMA.
it's necessary to indicate volume and period.
then you transfer the first data word in auddat, which will start the process: the word will be read at the frequency indicated and an interruption of level 4 will be generated indicating that you can send the following word. if you dont grab it, the sound transmission will end, and you will be forced to restart the process.

advantages (summary):
1. makes it possible to use samples of which the length depends only on the memory size.
2. amplitude and frequency modulation by using only one channel.
3. easy realization of many effects (inversion of sound, scratching...)
4. mix in real time several signals: fake 4+ channels!
to mix 2 sounds, it's enough to add the current amplitudes of the two signals by taking the differences of volume and frequency into account.
5. data dont need more to be in CHIP since the DMA is not used.

this mode is needed for the sampling.

last, the filter:
it cuts all the high frequencies, no more risks of aliasing.[...]
it's on bit 1 of port A of the CIA-A ($$bfe001). On old Amiga, it on/off the power-led.

ex.: eor.w #2,$bfe001

advices:
try echo effect by desynchronizing slightly two voices playing the same sound (and it's better if on the same stereo channel)
for good quality: use all the band of amplitude (+127 to -128), frequency not too low (min 7000), and for the loop, same value for the beginning and the end of the sample."

The End

credits: Philippe Rivaillon in "La Voix de Paula" - ANT issues 29/30 jan/feb 1992