approxRational, applied to two real fractional numbers x and epsilon,
returns the simplest rational number within epsilon of x.
A rational number y is said to be simpler than another y' if
chop granularises every sample in place as it is played, turning a
pattern of samples into a pattern of sample parts. Can be used to explore
granular synthesis.
Use an integer value to specify how many granules each sample is chopped into:
d1 $ chop 16 $ sound "arpy arp feel*4 arpy*4"
Different values of chop can yield very different results, depending on the
samples used:
You can also use chop (or striate) with very long samples to cut them into short
chunks and pattern those chunks. The following cuts a sample into 32 parts, and
plays it over 8 cycles:
d1 $ loopAt 8 $ chop 32 $ sound "bev"
The loopAt takes care of changing the speed of sample playback so that the
sample fits in the given number of cycles perfectly. As a result, in the above
the granules line up perfectly, so you can’t really hear that the sample has
been cut into bits. Again, this becomes more apparent when you do further
manipulations of the pattern, for example rev to reverse the order of the cut
up bits:
Applies a type of delay to a pattern.
It has three parameters, which could be called depth, time and feedback.
depth is and integer, and time and feedback are floating point numbers.
This adds a bit of echo:
d1 $ echo 4 0.2 0.5 $ sound "bd sn"
The above results in 4 echos, each one 50% quieter than the last, with 1/5th of a cycle between them.
echoWith is similar to echo, but instead of just decreasing volume to
produce echoes, echoWith applies a function each step and overlays the
result delayed by the given time.
d1 $ echoWith 2 "1%3" (# vowel "{a e i o u}%2") $ sound "bd sn"
In this case there are two _overlays_ delayed by 1/3 of a cycle, where each
has the vowel filter applied.
gap is similar to chop in that it granualizes every sample in place as it is played,
but every other grain is silent. Use an integer value to specify how many granules
each sample is chopped into:
d1 $ gap 8 $ sound "jvbass"
d1 $ gap 16 $ sound "[jvbass drum:4]"
hurry is similiar to fast in that it speeds up a pattern, but it also
increases the speed control by the same factor. So, if you’re triggering
samples, the sound gets higher in pitch. For example:
(A function that takes two ControlPatterns, and blends them together into
a new ControlPattern. An ControlPattern is basically a pattern of messages to
a synthesiser.)
Shifts between the two given patterns, using distortion.
loopAt makes a sample fit the given number of cycles. Internally, it
works by setting the unit parameter to "c", changing the playback
speed of the sample with the speed parameter, and setting setting
the density of the pattern to match.
d1 $ loopAt 4 $ sound "breaks125"
It’s a good idea to use this in conjuction with chop, so the break is chopped
into pieces and you don’t have to wait for the whole sample to start/stop.
d1 $ loopAt 4 $ chop 32 $ sound "breaks125"
Like all Tidal functions, you can mess about with this considerably. The below
example shows how you can supply a pattern of cycle counts to loopAt:
(Alias mt) Mod trigger. Aligns the start of a pattern to the
next cycle boundary where the cycle is evenly divisible by a given
number. qtrigger is equivalent to mtrigger 1.
In the following example, when activating the d1 pattern, it will start at the
same time as the next clap, even if it has to wait for 3 cycles. Once activated,
the arpy sound will play on every cycle, just like any other pattern:
(Alias qt) Quantise trigger. Aligns the start of the pattern
with the next cycle boundary. For example, this pattern will fade in
starting with the next cycle after the pattern is evaluated:
d1 $ qtrigger $ s "hh(5, 8)" # amp envL
Note that the pattern will start playing immediately. The start of the
pattern aligns with the next cycle boundary, but events will play before
if the pattern has events at negative timestamps (which most loops do).
These events can be filtered out, for example:
d1 $ qtrigger $ filterWhen (>= 0) $ s "hh(5, 8)"
Alternatively, you can use wait to achieve the same result:
slice is similar to chop and striate, in that it’s used to slice
samples up into bits. The difference is that it allows you to rearrange those
bits as a pattern.
smash is a combination of spread and striate — it cuts the samples
into the given number of bits, and then cuts between playing the loop
at different speeds according to the values in the list. So this:
spin will "spin" and layer up a pattern the given number of times,
with each successive layer offset in time by an additional 1/n of a cycle,
and panned by an additional 1/n. The result is a pattern that seems to spin
around. This function work well on multichannel systems.
Striate is a kind of granulator, cutting samples into bits in a similar to
chop, but the resulting bits are organised differently. For example:
d1 $ striate 3 $ sound "ho ho:2 ho:3 hc"
This plays the loop the given number of times, but triggers progressive portions
of each sample. So in this case it plays the loop three times, the first
time playing the first third of each sample, then the second time playing the
second third of each sample, and lastly playing the last third of each sample.
Replacing striate with chop above, one can hear that the 'chop version
plays the bits from each chopped-up sample in turn, while striate "interlaces"
the cut up bits of samples together.
You can also use striate with very long samples, to cut them into short
chunks and pattern those chunks. This is where things get towards granular
synthesis. The following cuts a sample into 128 parts, plays it over 8 cycles
and manipulates those parts by reversing and rotating the loops:
The striateBy function is a variant of striate with an extra
parameter which specifies the length of each part. The striateBy
function still scans across the sample over a single cycle, but if
each bit is longer, it creates a sort of stuttering effect. For
example the following will cut the bev sample into 32 parts, but each
will be 1/16th of a sample long:
d1 $ slow 32 $ striateBy 32 (1/16) $ sound "bev"
Note that striate and striateBy use the begin and end parameters
internally. This means that you probably shouldn't also specify begin or
end.
Align the start of a pattern with the time a pattern is evaluated,
rather than the global start time. Because of this, the pattern will
probably not be aligned to the pattern grid.
This aligns the start of a pattern to some value relative to the
time the pattern is evaluated. The provided function maps the evaluation
time (on the global cycle clock) to a new time, and then triggerWith
aligns the pattern's start to the time that's returned.
This is a more flexible triggering function. In fact, all the other trigger
functions are defined based on triggerWith. For example, trigger is just
triggerWith id.
In the next example, use d1 as a metronome, and play with different values
(from 0 to 1) on the const expression. You’ll notice how the clap is
displaced from the beginning of each cycle to the end, as the number increases:
d1 $ s "bd hh!3"
d2 $ triggerWith (const 0.1) $ s "clap"
weave applies one control pattern to a list of other control patterns, with
a successive time offset. It uses an OscPattern to apply the function at
different levels to each pattern, creating a weaving effect. For example:
In the above, the pan sine control pattern is slowed down by the given
number of cycles, in particular 16, and applied to all of the given sound
patterns. What makes this interesting is that the pan control pattern is
successively offset for each of the given sound patterns; because the pan is
closed down by 16 cycles, and there are four patterns, they are ‘spread out’,
i.e. with a gap of four cycles. For this reason, the four patterns seem to
chase after each other around the stereo field. Try listening on headphones to
hear this more clearly.
You can even have it the other way round, and have the effect parameters chasing
after each other around a sound parameter, like this:
d1 $ weave 16 (sound "arpy" >| n (run 8))
[ vowel "a e i"
, vowel "i [i o] o u"
, vowel "[e o]/3 [i o u]/2"
, speed "1 2 3"
]
Like append, but for a list of patterns. Interlaces them, playing the
first cycle from each in turn, then the second cycle from each, and so on. It
concatenates a list of patterns into a new pattern; each pattern in the list
will maintain its original duration. For example:
compress takes a pattern and squeezes it within the specified time span (i.e.
the ‘arc’). The new resulting pattern is a sped up version of the original.
d1 $ compress (1/4, 3/4) $ s "[bd sn]!"
In the above example, the pattern will play in an arc spanning from 25% to 75%
of the duration of a cycle. It is equivalent to:
d1 $ s "~ [bd sn]! ~"
Another example, where all events are different:
d1 $ compress (1/4, 3/4) $ n (run 4) # s "arpy"
It differs from zoom in that it preserves the original pattern but it speeds
up its events so to match with the new time period.
cosine - unipolar cosine wave. A pattern of continuous values
following a cosine with frequency of one cycle, and amplitude from
0 to 1. Equivalent to 0.25 ~> sine.
cosine2 - bipolar cosine wave. A pattern of continuous values
following a cosine with frequency of one cycle, and amplitude from
-1 to 1. Equivalent to 0.25 ~> sine2.
envL is a Pattern of continuous Double values, representing
a linear interpolation between 0 and 1 during the first cycle, then
staying constant at 1 for all following cycles. Possibly only
useful if you're using something like the retrig function defined
in tidal.el.
Functions which work on other functions (higher order functions)
every n f p applies the function f to p, but only affects
every n cycles.
It takes three inputs: how often the function should be applied (e.g. 3 to
apply it every 3 cycles), the function to be applied, and the pattern you are
applying it to. For example: to reverse a pattern every three cycles (and for
the other two play it normally)
d1 $ every 3 rev $ n "0 1 [~ 2] 3" # sound "arpy"
Note that if the function you’re applying requires additional parameters
itself (such as fast 2 to make a pattern twice as fast), then you’ll need to
wrap it in parenthesis, like so:
d1 $ every 3 (fast 2) $ n "0 1 [~ 2] 3" # sound "arpy"
Otherwise, the every function will think it is being passed too many parameters.
every' n o f p is like every n f p but with an offset of o cycles.
For example, every' 3 0 (fast 2) will speed up the cycle on cycles 0,3,6,…
whereas every' 3 1 (fast 2) will transform the pattern on cycles 1,4,7,….
With this in mind, setting the second argument of every' to 0 gives the
equivalent every function. For example, every 3 is equivalent to every' 3 0.
The every functions can be used to silence a full cycle or part of a cycle
by using silent or mask "~". Mask provides additional flexibility to turn on/off
individual steps.
d1 $ every 3 silent $ n "2 9 11 2" # s "hh27"
d1 $ every 3 (mask "~") $ n "2 9 10 2" # s "hh27"
d1 $ every 3 (mask "0 0 0 0") $ n "2 9 11 2" # s "hh27"
fastGap is similar to fast but maintains its cyclic alignment, i.e.,
rather than playing the pattern multiple times, it instead leaves a gap in
the remaining space of the cycle. For example, fastGap 2 p would squash the
events in pattern p into the first half of each cycle (and the second halves
would be empty). The factor should be at least 1.
foldEvery ns f p applies the function f to p, and is applied for
each cycle in ns.
It is similar to chaining multiple every functions together. It transforms
a pattern with a function, once per any of the given number of cycles. If a
particular cycle is the start of more than one of the given cycle periods, then
it it applied more than once.
d1 $ foldEvery [5,3] (|+ n 1) $ s "moog" # legato 1
The first moog samples are tuned to C2, C3 and C4. Note how on cycles that are
multiples of 3 or 5 the pitch is an octave higher, and on multiples of 15 the
pitch is two octaves higher, as the transformation is applied twice.
'fromMaybes; is similar to fromList, but allows values to
be optional using the Maybe type, so that Nothing results in
gaps in the pattern.
The following are equivalent:
> d1 $ n (fromMaybes [Just 0, Nothing, Just 2]) # s "superpiano"
> d1 $ n "0 ~ 2" # s "superpiano"
Slow down a pattern by the factors in the given time pattern, "squeezing"
the pattern to fit the slot given in the time pattern. It is the slow analogue
to fastSqueeze.
If the time pattern only has a single value in a cycle, slowSqueeze becomes equivalent to slow. These are equivalent:
d1 $ slow "<2 4>" $ s "bd*8"
d1 $ slowSqueeze "<2 4>" $ s "bd*8"
When the time pattern has multiple values, however, the behavior is a little
different. Instead, a slowed version of the pattern will be made for each value
in the time pattern, and they’re all combined together in a cycle according to
the structure of the time pattern. For example, these are equivalent:
d1 $ slowSqueeze "2 4 8 16" $ s "bd*8"
d1 $ s "bd*4 bd*2 bd bd/2"
as are these:
d1 $ slowSqueeze "2 4 [8 16]" $ s "bd*8"
d1 $ s "bd*4 bd*2 [bd bd/2]"
square - unipolar square wave. A pattern of continuous values
following a square wave with frequency of one cycle, and amplitude from
0 to 1.
| square is like sine, for square waves.
stack combines a list of Patterns into a new pattern, so that their
events are combined over time, i.e., all of the patterns in the list are played
simultaneously.
Similar to fastCat, but each pattern is given a relative duration.
You provide proportionate sizes of the patterns to each other for when they’re
concatenated into one cycle. The larger the value in the list, the larger
relative size the pattern takes in the final loop. If all values are equal
then this is equivalent to fastcat (e.g. the following two code fragments are
equivalent).
d1 $ fastcat [s "bd*4", s "hh27*8", s "superpiano" # n 0]
d1 $ timeCat [ (1, s "bd*4")
, (1, s "hh27*8")
, (1, s "superpiano" # n 0)
]
The given pattern transformation is applied only when the given test function
returns True. The test function will be called with the current cycle as
a number.
The above will only apply striate 4 to the pattern if the current
cycle number contains the number 4. So the fourth cycle will be
striated and the fourteenth and so on. Expect lots of striates after
cycle number 399.
Like when, but works on continuous time values rather than cycle numbers.
The following will apply # speed 2 only when the remainder of the current
Time divided by 2 is less than 0.5:
Plays a portion of a pattern, specified by a time arc (start and end time).
The new resulting pattern is played over the time period of the original pattern.
A pattern of numbers that speed up (or slow down) samples while they play.
In the following example, the sound starts at the original pitch and gets
higher as it plays:
d1 $ s "arpy" # accelerate 2
You can use a negative number to make the sound get lower. In this example, a
different acceleration is applied to each played note using state values:
A pattern of strings. When sent to SuperDirt, will be prepended to sample
folder names, separated by an underscore. This allows sample sets to be
organised into separate banks. See
https://github.com/musikinformatik/SuperDirt/pull/312
begin receives a pattern of numbers from 0 to 1 and skips the beginning
of each sample by the indicated proportion. I.e., 0 would play the sample from
the start, 1 would skip the whole sample, and 0.25 would cut off the first
quarter.
In this example, the first 3 ade samples are played on every cycle, but the
start point from which they are played changes on each cycle:
d1 $ n "0 1 2" # s "ade" # begin "<0 0.25 0.5 0.75>" # legato 1
In the style of classic drum-machines, cut will stop a playing sample as soon as another samples with in same cutgroup is to be played. An example would be an open hi-hat followed by a closed one, essentially muting the open.
Similar to begin, but cuts the end off samples, shortening them; e.g.
0.75 to cut off the last quarter of each sample.
d1 $ s "bev" >| begin 0.5 >| end "[0.65 0.55]"
The example above will play the sample two times for cycle, but the second time
will play a shorter segment than the first time, creating a kind of canon effect.
Used to control the amplitude (volume) of the sound. Values less than 1
make the sound quieter and values greater than 1 make the sound louder.
gain uses a power function, so the volume change around 1 is subtle, but it
gets more noticeable as it increases or decreases. Typical values for gain are
between 0 and 1.5.
a pattern of numbers to specify the hold time (in seconds) of an envelope applied to each sample. Only takes effect if attack and release are also specified.
A pattern of numbers. Specifies whether the pitch of played samples should be tuned relative to their pitch metadata, if it exists. When set to 1, pitch metadata is applied. When set to 0, pitch metadata is ignored.
a pattern of numbers. An "orbit" is a global parameter context for patterns. Patterns with the same orbit will share hardware output bus offset and global effects, e.g. reverb and delay. The maximum number of orbits is specified in the superdirt startup, numbers higher than maximum will wrap around.
A pattern of numbers which changes the speed of sample playback which also
changes pitch. Negative values will play the sample backwards.
d1 $ slow 5 $ s "sax:5" # legato 1 # speed 0.5
This will play the sax:5 sample at half its rate. As a result, the sample will
last twice the normal time, and will be pitched a whole octave lower. This is
equivalent to d1 $ slow 5 $ s "sax:5" # legato 1 |- note 12.
d1 $ fast 2 $ s "breaks125:1" # cps (125/60/4) # speed (-2)
In the above example, the break (which lasts for exactly one bar at 125 BPM), will be played backwards, and at double speed (so, we use fast 2 to fill the whole cycle).
A pattern of numbers that indicates the total duration of sample playback in seconds.
This sustain refers to the whole playback duration and is not to be confused with the sustain level of a typical ADSR envelope.
d1 $ fast 2 $ s "breaks125:1" # cps (120/60/4) # sustain 1
At 120 BPM, a cycle lasts for two seconds. In the above example, we cut the
sample so it plays just for one second, and repeat this part two times, so we
fill the whole cycle. Note that sample pitch isn’t modified.
d1 $ s "breaks125:2!3" # cps (120/60/4) # sustain "0.4 0.2 0.4" # begin "0 0 0.4"
Here, we take advantage that sustain receives a pattern to build a different
break from the original sample.
timescale is the main function used to activate time-stretching, and usually
the only one you need. It receives a single parameter which is the stretching
rate to apply.
You can use any positive number as the ratio, but the particular method used is
designed for ratios greater than 1, and work reasonably well for values between
0.1 and 3.
In the example above, we set tempo at 130 beats per minute. But we want to play
one of the breaks152 samples, which are, as indicated, at 152 BPM. So, the
ratio we want is 152 over 130. This will slow down the sample to fit in our 130
BPM tempo.
The algorithm used to time-stretch a sample divides a sample in many little parts, modifies them, and puts them all together again. It uses one particular parameter, called windowSize, which is the length of each sample part.
The windowSize value is automatically calculated, but can be changed with timescalewin. The windowSize value is multiplied by the number provided.
timescalewin can be used to improve the quality of time-stretching for some samples, or simply as an effect.
Consider the following two examples. In the first one, timescalewin 0.01 makes
the window size a lot smaller, and the extreme chopping of the sample causes
a rougher sound. In the second one, timescalewin 10 makes the chunks a lot
bigger. The method used overlaps the treated chunks when recomposing the sample,
and, with the bigger window size, this overlap is noticeable and causes a kind
of delay effect.
Used in conjunction with speed. It accepts values of r (rate, default
behavior), c (cycles), or s (seconds). Using unit "c" means speed
will be interpreted in units of cycles, e.g. speed "1" means samples will be
stretched to fill a cycle. Using unit "s" means the playback speed will be
adjusted so that the duration is the number of seconds specified by speed.
In the following example, speed 2 means that samples will be stretched to fill
half a cycle:
d1 $ stack [
s "sax:5" # legato 1 # speed 2 # unit "c",
s "bd*2"
]
For example, the following will play the sound pattern "bd sn kurt" twice as
fast (i.e., so it repeats twice per cycle), and the vowel pattern three times
as fast:
d1 $ sound (fast 2 "bd sn kurt")
# fast 3 (vowel "a e o")
The first parameter can be patterned to, for example, play the pattern at twice
the speed for the first half of each cycle and then four times the speed for the
second half:
fastSqueeze speeds up a pattern by a time pattern given as input,
squeezing the resulting pattern inside one cycle and playing the original
pattern at every repetition.
To better understand how it works, compare it with fast:
Example1 expression
>>> fast "1 2" $ s "bd sn"(0>½)|s: "bd"(½>¾)|s: "bd"(¾>1)|s: "sn"
This will give bd played in the first half cycle, and bd sn in the second
half. On the other hand, using fastSqueeze;
The original pattern will play in the first half, and two repetitions of the
original pattern will play in the second half. That is, every repetition
contains the whole pattern.
If the time pattern has a single value, it becomes equivalent to fast:
d1 $ fastSqueeze 2 $ s "bd sn"
d1 $ fast 2 $ s "bd sn"
d1 $ s "[bd sn]*2"
rev p returns p with the event positions in each cycle reversed (or
mirrored).
For example rev "1 [~ 2] ~ 3" is equivalent to rev "3 ~ [2 ~] 1".
Note that rev reverses on a cycle-by-cycle basis. This means that rev (slow
2 "1 2 3 4") would actually result in (slow 2 "2 1 4 3"). This is because the
slow 2 makes the repeating pattern last two cycles, each of which is reversed
independently.
In practice rev is generally used with conditionals, for example with every:
For example, the following will play the sound pattern "bd sn kurt" twice as
slow (i.e., so it repeats once every two cycles), and the vowel pattern three
times as slow:
d1 $ sound (slow 2 "bd sn kurt")
# slow 3 (vowel "a e o")
Splits queries that span cycles. For example `query p (0.5, 1.5)` would be
turned into two queries, `(0.5,1)` and `(1,1.5)`, and the results
combined. Being able to assume queries don't span cycles often
makes transformations easier to specify.
Like unwrap, but cycles of the inner patterns are compressed to fit the
timespan of the outer whole (or the original query if it's a continuous pattern?)
TODO - what if a continuous pattern contains a discrete one, or vice-versa?
TODO - steps
Turns a pattern of patterns into a single pattern.
(this is actually join)
1/ For query arc, get the events from the outer pattern pp
2/ Query the inner pattern using the part of the outer
3/ For each inner event, set the whole and part to be the intersection
of the outer whole and part, respectively
4 Concatenate all the events together (discarding wholesparts that didn't intersect)
TODO - what if a continuous pattern contains a discrete one, or vice-versa?
The above takes the standard scaleTable as a starting point and adds two custom scales to it. You’ll be able to use the new function in place of the normal one:
Outputs a list of all available scales and their corresponding notes. For
example, its first entry is ("minPent",[0,3,5,7,10]) which means that
a minor pentatonic scale is formed by the root (0), the minor third (3 semitones
above the root), the perfect fourth (5 semitones above the root), etc.
As the list is big, you can use the Haskell function lookup to look up a
specific scale: lookup "phrygian" scaleTable. This will output
Just [0.0,1.0,3.0,5.0,7.0,8.0,10.0].
You can also do a reverse lookup into the scale table. For example:
filter ( \(_, x) -> take 3 x == [0,2,4] ) scaleTable
The above example will output all scales of which the first three notes are
the root, the major second (2 semitones above the fundamental), and the major
third (4 semitones above the root).
Shifts an Arc to one of equal duration that starts within cycle zero.
(Note that the output Arc probably does not start *at* Time 0 --
that only happens when the input Arc starts at an integral Time.)
This provides exactly the same information as cyclesInArc,
except that this represents its output as Arcs,
whereas cyclesInArc represents the same information as integral indices.
(The Arc from 0 to 1 corresponds to the index 0,
the one from 1 to 2 has index 1, etc.)
Returns the numbers of the cycles that the input Arc overlaps
(excluding the input Arc's endpoint, unless it has duration 0 --
see "Edge cases" below).
(The "cycle number" of an Arc is equal to its start value.
Thus, for instance, cyclesInArc (Arc 0 1.5) == [0,1].)
Edge cases:
> cyclesInArc $ Arc 0 1.0001 == [0,1]
> cyclesInArc $ Arc 0 1 == [0] -- the endpoint is excluded
> cyclesInArc $ Arc 1 1 == [1] -- unless the Arc has duration 0
PITFALL: Don't be fooled by the name. The output cycles
are not necessarily completely contained in the input Arc,
but they definitely overlap it,
and they include every cycle that overlaps it.
subArc i j is the timespan that is the intersection of i and j.
intersection
The definition is a bit fiddly as results might be zero-width, but
not at the end of an non-zero-width arc - e.g. (0,1) and (1,2) do
not intersect, but (1,1) (1,1) does.
arpeggiate finds events that share the same timespan, and spreads
them out during that timespan, so for example arpeggiate "[bd,sn]"
gets turned into "bd sn". Useful for creating arpeggios/broken chords.
beat structures a pattern by picking subdivisions of a cycle.
Takes in a pattern that tells it which parts to play (polyphony is recommeded here),
and the number of parts by which to subdivide the cycle (also pattern-able).
For example:
> d1 $ beat "[3,4.2,9,11,14]" 16 $ s "sd"
bite n ipat pat slices a pattern pat into n pieces, then uses the
ipat pattern of integers to index into those slices. So bite 4 "0 2*2" (run
8) is the same as "[0 1] [4 5]*2".
I.e., it allows you to slice each cycle into a given number of equal sized
bits, and then pattern those bits by number. It’s similar to slice, but is
for slicing up patterns, rather than samples. The following slices the pattern
into four bits, and then plays those bits in turn:
d1 $ bite 4 "0 1 2 3" $ n "0 .. 7" # sound "arpy"
Of course that doesn’t actually change anything, but then you can reorder those bits:
d1 $ bite 4 "2 0 1 3" $ n "0 .. 7" # sound "arpy"
The slices bits of pattern will be squeezed or contracted to fit:
brak makes a pattern sound a bit like a breakbeat. It does this by, every
other cycle, squashing the pattern to fit half a cycle, and offsetting it by a
quarter of a cycle.
d1 $ sound (brak "bd sn kurt")
d1 $ brak $ sound "[feel feel:3, hc:3 hc:2 hc:4 ho:1]"
plays a melody randomly choosing one of the four notes "a", "e", "g", "c".
As with all continuous patterns, you have to be careful to give them structure; in this case choose gives you an infinitely detailed stream of random choices.
Given a pattern of doubles, chooseBy normalizes them so that each
corresponds to an index in the provided list. The returned pattern
contains the corresponding elements in the list.
It is like choose, but instead of selecting elements of the list randomly, it
uses the given pattern to select elements.
The first argument indicates the (patternable) number of notes to insert,
and the second argument is the base pattern of "anchor notes" that gets transformed.
Treats the given pattern p as having n chunks, and applies the function f to one of those sections per cycle.
Running:
- from left to right if chunk number is positive
- from right to left if chunk number is negative
contrast f f' p p' splits the control pattern p' in two, applying
the function f to one and f' to the other. This depends on
whether events in p' contain values matching with those in p.
For example, in
contrast (# crush 3) (# vowel "a") (n "1") $ n "0 1" # s "bd sn" # speed 3
the first event will have the vowel effect applied and the second will have
the crush applied.
contrast is like an if-else-statement over patterns. For contrast t f p
you can think of t as the true branch, f as the false branch, and p as
the test.
You can use any control pattern as a test of equality, e.g., n "1", speed
"0.5", or things like that. This lets you choose specific properties of the
pattern you’re transforming for testing, like in the following example,
d1 $ contrast (|+ n 12) (|- n 12) (n "c") $ n (run 4) # s "superpiano"
where every note that isn’t middle-c will be shifted down an octave but
middle-c will be shifted up to c5.
Since the test given to contrast is also a pattern, you can do things like have
it alternate between options:
d1 $ contrast (|+ n 12) (|- n 12) (s "<superpiano superchip>")
$ s "superpiano superchip" # n 0
If you listen to this you’ll hear that which instrument is shifted up and which
instrument is shifted down alternates between cycles.
contrastBy is contrastBy is the general version of contrast, in which you can specify an abritrary boolean function that will be used to compare the control patterns.
d2 $ contrastBy (>=) (|+ n 12) (|- n 12) (n "2") $ n "0 1 2 [3 4]" # s "superpiano"
You can use the euclid function to apply a Euclidean algorithm over a
complex pattern, although the structure of that pattern will be lost:
d1 $ euclid 3 8 $ sound "bd*2 [sn cp]"
In the above, three sounds are picked from the pattern on the right according
to the structure given by the euclid 3 8. It ends up picking two bd sounds, a
cp and missing the sn entirely.
A negative first argument provides the inverse of the euclidean pattern.
These types of sequences use "Bjorklund's algorithm", which wasn't made for
music but for an application in nuclear physics, which is exciting. More
exciting still is that it is very similar in structure to the one of the first
known algorithms written in Euclid's book of elements in 300 BC. You can read
more about this in the paper
The Euclidean Algorithm Generates Traditional Musical Rhythms
by Toussaint. Some examples from this paper are included below,
including rotation as a third parameter in some cases (see euclidOff).
Pattern
Example
(2,5)
A thirteenth century Persian rhythm called Khafif-e-ramal.
(3,4)
The archetypal pattern of the Cumbia from Colombia, as well as
a Calypso rhythm from Trinidad.
(3,5,2)
Another thirteenth century Persian rhythm by the name of
Khafif-e-ramal, as well as a Rumanian folk-dance rhythm.
(3,7)
A Ruchenitza rhythm used in a Bulgarian folk-dance.
(3,8)
The Cuban tresillo pattern.
(4,7)
Another Ruchenitza Bulgarian folk-dance rhythm.
(4,9)
The Aksak rhythm of Turkey.
(4,11)
The metric pattern used by Frank Zappa in his piece titled
Outside Now.
(5,6)
Yields the York-Samai pattern, a popular Arab rhythm.
(5,7)
The Nawakhat pattern, another popular Arab rhythm.
(5,8)
The Cuban cinquillo pattern.
(5,9)
A popular Arab rhythm called Agsag-Samai.
(5,11)
The metric pattern used by Moussorgsky in
Pictures at an Exhibition.
(5,12)
The Venda clapping pattern of a South African children’s song.
(5,16)
The Bossa-Nova rhythm necklace of Brazil.
(7,8)
A typical rhythm played on the Bendir (frame drum).
(7,12)
A common West African bell pattern.
(7,16,14)
A Samba rhythm necklace from Brazil.
(9,16)
A rhythm necklace used in the Central African Republic.
(11,24,14)
A rhythm necklace of the Aka Pygmies of Central Africa.
(13,24,5)
Another rhythm necklace of the Aka Pygmies of the upper Sangha.
There was once a shorter alias e for this function. It has been removed, but you
may see references to it in older Tidal code.
euclidFull n k pa pb stacks euclid n k pa with euclidInv n k pb. That
is, it plays one pattern on the euclidean rhythm and a different pattern on
the off-beat.
For example, to implement the traditional flamenco rhythm, you could use hard
claps for the former and soft claps for the latter:
d1 $ euclidFull 3 7 "realclaps" ("realclaps" # gain 0.8)
fastspread works the same as spread, but the result is squashed into a single cycle. If you gave four values to spread, then the result would seem to speed up by a factor of four. Compare these two:
The fit function takes a pattern of integer numbers, which are used to select values from the given list. What makes this a bit strange is that only a given number of values are selected each cycle. For example:
The above fits three samples into the pattern, i.e. for the first cycle this
will be "bd", "sn" and "arpy", giving the result "bd [~ sn] arpy sn"
(note that we start counting at zero, so that 0 picks the first value). The
following cycle the next three values in the list will be picked, i.e.
"arpy:1", "casio" and "bd", giving the pattern
"arpy:1 [~ casio] bd casio" (note that the list wraps round here).
fit' is a generalization of fit, where the list is instead constructed
by using another integer pattern to slice up a given pattern. The first argument
is the number of cycles of that latter pattern to use when slicing. It's easier
to understand this with a few examples:
d1 $ sound (fit' 1 2 "0 1" "1 0" "bd sn")
So what does this do? The first 1 just tells it to slice up a single cycle of
"bd sn". The 2 tells it to select two values each cycle, just like the first
argument to fit. The next pattern "0 1" is the "from" pattern which tells
it how to slice, which in this case means "0" maps to "bd", and "1" maps
to "sn". The next pattern "1 0" is the "to" pattern, which tells it how to
rearrange those slices. So the final result is the pattern "sn bd".
which uses chop to break a single sample into individual pieces, which fit' then puts into a list (using the run 4 pattern) and reassembles according to the complicated integer pattern.
The fix function applies another function to matching events in a pattern of
controls. fix is contrast where the false-branching function is set to the
identity id. It is like contrast, but one function is given and applied to
events with matching controls.
For example, the following only adds the crush control when the n control
is set to either 1 or 4:
You can be quite specific; for example, the following applies the function
hurry 2 to sample 1 of the drum sample set, and leaves the rest as they are:
The fixRange function isn’t very user-friendly at the moment, but you can
create a fix variant with a range condition. Any value of a ControlPattern
wich matches the values will apply the passed function.
flatpat takes a Pattern of lists and pulls the list elements as
separate Events. For example, the following code uses flatpat in combination with listToPat to create an alternating pattern of chords:
d1 $ n (flatpat $ listToPat [[0,4,7],[(-12),(-8),(-5)]])
# s "superpiano" # sustain 2
This code is equivalent to:
d1 $ n ("[0,4,7] [-12,-8,-5]") # s "superpiano" # sustain 2
Like ghost', but a user-supplied function describes how to alter the pattern.
In this example, ghost notes are applied to the snare hit, but these notes will
be louder, not quieter, and the sample will have its beginning slightly cut:
d1 $ slow 2
$ ghostWith (1/16) ((|*| gain 1.1) . (|> begin 0.05))
$ sound "sn"
Given a start point and a duration (both specified in cycles), this
generates a control pattern that makes a sound begin at the start
point and last the duration.
The following are equivalent:
d1 $ slow 2 $ s "bev" # grain 0.2 0.1 # legato 1
d1 $ slow 2 $ s "bev" # begin 0.2 # end 0.3 # legato 1
grain is defined as:
grain s d = 'Sound.Tidal.Params.begin' s # 'Sound.Tidal.Params.end' (s+d)
This will apply striate 4 for every even cycle and apply # coarse "24 48" for every odd.
Detail: As you can see the test function is arbitrary and does not rely on
anything Tidal specific. In fact it uses only plain Haskell functionality, that
is: it calculates the modulo of 2 of the current cycle which is either 0 (for
even cycles) or 1. It then compares this value against 0 and returns the result,
which is either True or False. This is what the ifp signature's first part
signifies: (Int -> Bool), a function that takes a whole number and returns
either True or False.
A simpler version of ur that just provides name-value bindings that are
reflected in the provided pattern.
inhabit allows you to link patterns to some String, or in other words,
to give patterns a name and then call them from within another pattern of
Strings.
For example, we can make our own bassdrum, hi-hat and snaredrum kit:
do
let drum = inhabit [ ("bd", s "sine" |- accelerate 1.5)
, ("hh", s "alphabet:7" # begin 0.7 # hpf 7000)
, ("sd", s "invaders:3" # speed 12)
]
d1 $ drum "[bd*8?, [~hh]*4, sd(6,16)]"
inhabit can be very useful when using MIDI controlled drum machines, since you
can give understandable drum names to patterns of notes.
inside carries out an operation inside a cycle.
For example, while rev "0 1 2 3 4 5 6 7" is the same as "7 6 5 4 3 2 1 0",
inside 2 rev "0 1 2 3 4 5 6 7" gives "3 2 1 0 7 6 5 4".
What this function is really doing is ‘slowing down’ the pattern by a given
factor, applying the given function to it, and then ‘speeding it up’ by the same
factor. In other words, this:
inside 2 rev "0 1 2 3 4 5 6 7"
Is doing this:
fast 2 $ rev $ slow 2 "0 1 2 3 4 5 6 7"
so rather than whole cycles, each half of a cycle is reversed.
Just like rand but for whole numbers, irand n generates a pattern of (pseudo-) random whole numbers between 0 to n-1 inclusive. Notably used to pick a random
samples from a folder:
Divides a pattern into a given number of subdivisions, plays the subdivisions
in order, but increments the starting subdivision each cycle. The pattern
wraps to the first subdivision after the last subdivision is played.
Example:
d1 $ iter 4 $ sound "bd hh sn cp"
This will produce the following over four cycles:
bd hh sn cp
hh sn cp bd
sn cp bd hh
cp bd hh sn
There is also iter', which shifts the pattern in the opposite direction.
The jux function creates strange stereo effects by applying a
function to a pattern, but only in the right-hand channel. For
example, the following reverses the pattern on the righthand side:
When passing pattern transforms to functions like jux and every,
it's possible to chain multiple transforms together with . (function
composition). For example this both reverses and halves the playback speed of
the pattern in the righthand channel:
In addition to jux, jux' allows using a list of pattern
transformations. Resulting patterns from each transformation will be spread via
pan from left to right.
For example, the following will put iter 4 of the pattern to the far left
and palindrome to the far right. In the center, the original pattern will
play and the chopped and the reversed version will appear mid left and mid
right respectively.
With jux, the original and effected versions of the pattern are
panned hard left and right (i.e., panned at 0 and 1). This can be a
bit much, especially when listening on headphones. The variant juxBy
has an additional parameter, which brings the channel closer to the
centre. For example:
d1 $ juxBy 0.5 (fast 2) $ sound "bd sn:1"
In the above, the two versions of the pattern would be panned at 0.25
and 0.75, rather than 0 and 1.
layer takes a list of Pattern-returning functions and a seed element,
stacking the result of applying the seed element to each function in the list.
It allows you to layer up multiple functions on one pattern. For example, the following
will play two versions of the pattern at the same time, one reversed and one at twice
the speed:
Returns the nth iteration of a
Lindenmayer System
with given start sequence.
It takes an integer b, a Lindenmayer system rule set, and an initiating
string as input in order to generate an L-system tree string of b iterations.
It can be used in conjunction with a step function to convert the generated
string into a playable pattern. For example,
generates an L-system with initiating string "0" and maps it onto a list
of samples.
Complex L-system trees with many rules and iterations can sometimes result in unwieldy strings. Using take n to only use the first n elements of the string, along with a slow function, can make the generated values more manageable.
lindenmayerI converts the resulting string into a a list of integers
with fromIntegral applied (so they can be used seamlessly where floats or
rationals are required)
linger is similar to trunc, in that it truncates a pattern so that
only the first fraction of the pattern is played, but the truncated part of the
pattern loops to fill the remainder of the cycle.
markovPat n xi tp generates a one-cycle pattern of n steps in a Markov
chain starting from state xi with transition matrix tp. Each row of the
transition matrix is automatically normalized. For example:
mask takes a boolean pattern and ‘masks’ another pattern with it. That is,
events are only carried over if they match within a ‘true’ event in the binary
pattern, i.e., it removes events from the second pattern that don't start during
an event from the first.
For example, consider this kind of messy rhythm without any rests.
Due to the use of slowcat here, the same mask is first applied to "sn*8" and
in the next cycle to "[cp*4 bd*4, hc*5]".
You could achieve the same effect by adding rests within the slowcat patterns,
but mask allows you to do this more easily. It kind of keeps the rhythmic
structure and you can change the used samples independently, e.g.,
For specifying a boolean pattern according to a list of offsets
(aka inter-onset intervals). For example necklace 12 [4,2] is
the same as "t f f f t f t f f f t f". That is, 12 steps per cycle,
with true values alternating between every 4 and every 2 steps.
Generates values in [0,1] that follows a normal (bell-curve) distribution.
One possible application is to "humanize" drums with a slight random delay:
d1 $
s "bd sn bd sn"
# nudge (segment 4 (0.01 * normal))
Implemented with the Box-Muller transform.
* the max ensures we don't calculate log 0
* the rot in u2 ensures we don't just get the same value as u1
* clamp the Box-Muller generated values in a [-3,3] range
off is similar to superimpose, in that it applies a function to a pattern
and layers up the results on top of the original pattern. The difference
is that off takes an extra pattern being a time (in cycles) to shift the
transformed version of the pattern by.
The following plays a pattern on top of itself, but offset by an eighth of a
cycle, with a distorting bitcrush effect applied:
outside is the inverse of the inside function. outside applies its function outside the cycle.
Say you have a pattern that takes 4 cycles to repeat and apply the rev function:
d1 $ rev $ cat [s "bd bd sn",s "sn sn bd", s"lt lt sd", s "sd sd bd"]
The above generates:
d1 $ rev $ cat [s "sn bd bd",s "bd sn sn", s "sd lt lt", s "bd sd sd"]
d1 $ rev $ cat [s "bd sd sd", s "sd lt lt", s "sn sn bd", s "bd bd sn"]
Notice that the whole idea has been reversed. What this function is really doing
is ‘speeding up’ the pattern by a given factor, applying the given function to
it, and then ‘slowing it down’ by the same factor. In other words, this:
d1 $ slow 4 $ rev $ fast 4
$ cat [s "bd bd sn",s "sn sn bd", s"lt lt sd", s "sd sd bd"]
This compresses the idea into a single cycle before rev operates and then slows it back to the original speed.
The perlin function produces a new random value to move to every cycle. If
you want a new random value to be generated more or less frequently, you can use
fast or slow, respectively:
The above will generate a smooth random cutoff pattern that repeats every cycle
without any reversals or discontinuities (because the 2D path is a circle).
See also: perlin2, which only needs one input because it uses the cycle count
as the second input.
1D Perlin (smooth) noise, works like rand but smoothly moves between random
values each cycle. perlinWith takes a pattern as the random number generator's
"input" instead of automatically using the cycle count.
Given a sample's directory name and number, this generates a string
suitable to pass to fromString to create a 'Pattern String'.
samples is a Pattern-compatible interface to this function.
Syncopates a rhythm, shifting (delaying) each event halfway into its arc
(timespan).
In mini-notation terms, it basically turns every instance of a into [~ a],
e.g., "a b [c d] e" becomes the equivalent of
"[~ a] [~ b] [[~ c] [~ d]] [~ e]".
Every beat then becomes an offbeat, and so the overall effect is to
syncopate a pattern.
In the following example, you can hear that the piano chords play between the
snare and the bass drum. In 4/4 time, they are playing in the 2 and a half,
and 4 and a half beats:
do
resetCycles
d1 $ stack [
press $ n "~ c'maj ~ c'maj" # s "superpiano" # gain 0.9 # pan 0.6,
s "[bd,clap sd bd sd]" # pan 0.4
] # cps (90/60/4)
In the next example, the C major chord plays before the G major. As the slot
that occupies the C chord is that of one eighth note, it is displaced by press
only a sixteenth note:
do
resetCycles
d1 $ stack [
press $ n "~ [c'maj ~] ~ ~" # s "superpiano" # gain 0.9 # pan 0.6,
press $ n "~ g'maj ~ ~" # s "superpiano" # gain 0.9 # pan 0.4,
s "[bd,clap sd bd sd]"
] # cps (90/60/4)
quantise limits values in a Pattern (or other Functor) to n equally spaced
divisions of 1.
It is useful for rounding a collection of numbers to some particular base
fraction. For example,
quantise 5 [0, 1.3 ,2.6,3.2,4.7,5]
It will round all the values to the nearest (1/5)=0.2 and thus will output
the list [0.0,1.2,2.6,3.2,4.8,5.0]. You can use this function to force a
continuous pattern like sine into specific values. In the following example:
d1 $ s "superchip*8" # n (quantise 1 $ range (-10) (10) $ slow 8 $ cosine)
# release (quantise 5 $ slow 8 $ sine + 0.1)
all the releases selected be rounded to the nearest 0.1 and the notes selected
to the nearest 1.
quantise with fractional inputs does the consistent thing: quantise 0.5
rounds values to the nearest 2, quantise 0.25 rounds the nearest 4, etc.
rand is an oscillator that generates a continuous pattern of (pseudo-)random
numbers between 0 and 1.
For example, to randomly pan around the stereo field:
d1 $ sound "bd*8" # pan rand
Or to enjoy a randomised speed from 0.5 to 1.5, add 0.5 to it:
d1 $ sound "arpy*4" # speed (rand + 0.5)
To make the snares randomly loud and quiet:
sound "sn sn ~ sn" # gain rand
Numbers coming from this pattern are 'seeded' by time. So if you reset time
(using resetCycles, setCycle, or cps) the random pattern will emit the
exact same _random_ numbers again.
In cases where you need two different random patterns, you can shift
one of them around to change the time from which the _random_ pattern
is read, note the difference:
jux (# gain rand) $ sound "sn sn ~ sn" # gain rand
and with the juxed version shifted backwards for 1024 cycles:
jux (# ((1024 <~) $ gain rand)) $ sound "sn sn ~ sn" # gain rand
range will take a pattern which goes from 0 to 1 (like sine), and range it to a different range - between the first and second arguments. In the below example, `range 1 1.5` shifts the range of sine1 from 0 - 1 to 1 - 1.5.
rangex is an exponential version of range, good for using with
frequencies. For example, range 20 2000 "0.5" will give 1010 - halfway
between 20 and 2000. But rangex 20 2000 0.5 will give 200 - halfway
between on a logarithmic scale. This usually sounds better if you’re using the
numbers as pitch frequencies. Since rangex uses logarithms, don’t try to scale
things to zero or less.
rolled plays each note of a chord quickly in order, as opposed to
simultaneously; to give a chord a harp-like or strum effect.
Notes are played low to high, and are evenly distributed within (14) of the chord event length, as opposed to arparpeggiate that spread the notes over the whole event.
rot n p "rotates" the values in a pattern p by n beats to the left,
preserving its structure. For example, in the following, each value will shift
to its neighbour's position one step to the left, so that b takes the place of
a, a of c, and c of b:
rot 1 "a ~ b c"
The result is equivalent of:
"b ~ c a"
The first parameter is the number of steps, and may be given as a pattern. For example, in
runMarkov n tmat xi seed generates a Markov chain (as a list) of length n
using the transition matrix tmat starting from initial state xi, starting
with random numbers generated from seed
Each entry in the chain is the index of state (starting from zero).
Each row of the matrix will be automatically normalized. For example:
runMarkov 8 [[2,3], [1,3]] 0 0
will produce a two-state chain 8 steps long, from initial state 0, where the
transition probability from state 0->0 is 25, 0->1 is 35, 1->0 is 1/4, and
1->1 is 3/4.
scramble n p is like shuffle but randomly selects from the parts
of p instead of making permutations.
For example, scramble 3 "a b c" will randomly select 3 parts from
"a""b" and "c", possibly repeating a single part.
This could also be called “sampling with replacement”.
segment n p ’samples’ the pattern p at a rate of n events per cycle.
Useful for turning a continuous pattern into a discrete one.
In the following example, the pattern originates from the shape of a sine
wave, a continuous pattern. Without segment, the samples will get triggered
at an undefined frequency which may be very high.
d1 $ n (slow 2 $ segment 16 $ range 0 32 $ sine) # sound "amencutup"
The function seqP allows you to define when
a sound within a list starts and ends. The code below contains three
separate patterns in a stack, but each has different start times
(zero cycles, eight cycles, and sixteen cycles, respectively). All
patterns stop after 128 cycles:
shuffle n p evenly divides one cycle of the pattern p into n parts,
and returns a random permutation of the parts each cycle. For example,
shuffle 3 "a b c" could return "a b c", "a c b", "b a c", "b c a",
"c a b", or "c b a". But it will never return "a a a", because that
is not a permutation of the parts.
This could also be called “sampling without replacement”.
slowstripe n p is the same as stripe, but the result is also
n times slower, so that the mean average duration of the stripes
is exactly one cycle, and every nth stripe starts on a cycle
boundary (in Indian classical terms, the sam).
smooth receives a pattern of numbers and linearly goes from one to the next, passing through all of them. As time is cycle-based, after reaching the last number in the pattern, it will smoothly go to the first one again.
snowball takes a function that can combine patterns (like +),
a function that transforms a pattern (like slow),
a depth, and a starting pattern,
it will then transform the pattern and combine it with the last transformation until the depth is reached.
This is like putting an effect (like a filter) in the feedback of a delay line; each echo is more affected.
Applies a function to a pattern and cats the resulting pattern, then continues
applying the function until the depth is reached this can be used to create
a pattern that wanders away from the original pattern by continually adding
random numbers.
As sometimesBy, but applies the given transformation to the pattern in its entirety
before filtering its actual appearances. Less efficient than sometimesBy but may
be useful when the passed pattern transformation depends on properties of the
pattern before probabilities are taken into account.
The spread function allows you to take a pattern transformation
which takes a parameter, such as slow, and provide several
parameters which are switched between. In other words it "spreads" a
function across several values.
Taking a simple high hat loop as an example:
d1 $ sound "ho ho:2 ho:3 hc"
We can slow it down by different amounts, such as by a half:
d1 $ slow 2 $ sound "ho ho:2 ho:3 hc"
Or by four thirds (i.e. speeding it up by a third; 4%3 means four over
three):
d1 $ slow (4%3) $ sound "ho ho:2 ho:3 hc"
But if we use spread, we can make a pattern which alternates between
the two speeds:
Note that if you pass ($) as the function to spread values over, you
can put functions as the list of values. (spreadf is an alias for spread ($).)
For example:
This is quite a messy area of Tidal—due to a slight difference of
implementation this sounds completely different! One advantage of
using spread' though is that you can provide polyphonic parameters, e.g.:
sseq acts as a kind of simple step-sequencer using strings. For example,
sseq "sn" "x x 12" is equivalent to the pattern of strings given by "sn ~
sn ~ sn:1 sn:2 ~". sseq substitutes the given string for each x, for each number
it substitutes the string followed by a colon and the number, and for everything
else it puts in a rest.
In other words, sseq generates a pattern of strings in exactly the syntax you’d want for selecting samples and that can be fed directly into the s function.
Uses the first (binary) pattern to switch between the following
two patterns. The resulting structure comes from the binary
pattern, not the source patterns. (In sew, by contrast, the resulting structure comes from the source patterns.)
The following uses a euclidean pattern to control CC0:
stretch takes a pattern, and if there’s silences at the start or end of the
current cycle, it will zoom in to avoid them. The following are equivalent:
stripe n p: repeats pattern pn times per cycle, i.e., the first
parameter gives the number of cycles to operate over. So, it is similar to
fast, but with random durations. For example stripe 2 will repeat a pattern
twice, over two cycles
In the following example, the start of every third repetition of the d1
pattern will match with the clap on the d2 pattern.
The repetitions will be contiguous (touching, but not overlapping) and the
durations will add up to a single cycle. n can be supplied as a pattern of
integers.
The structure comes from a boolean pattern, i.e. a binary pattern containing
true or false values. Above we only used true values, denoted by t. It’s also
possible to include false values with f, which struct will simply treat as
silence. For example, this would have the same outcome as the above:
d1 $ struct ("t f t*2 f") $ sound "cp"
These true / false binary patterns become useful when you conditionally
manipulate them, for example, ‘inverting’ the values using every and inv:
In the above, the boolean values will be ‘inverted’ every third cycle, so that
the structure comes from the fs rather than t. Note that euclidean patterns
also create true/false values, for example:
d1 $ struct (every 3 inv "t(3,8)") $ sound "cp"
In the above, the euclidean pattern creates "t f t f t f f t" which gets
inverted to "f t f t f t t f" every third cycle. Note that if you prefer you
can use 1 and 0 instead of t and f.
stutter n t pat repeats each event in patn times, separated by t time (in fractions of a cycle).
It is like echo that doesn't reduce the volume, or ply if you controlled the timing.
superimpose plays a modified version of a pattern at the same time as the
original pattern, resulting in two patterns being played at the same time. The
following are equivalent:
d1 $ superimpose (fast 2) $ sound "bd sn [cp ht] hh"
d1 $ stack [sound "bd sn [cp ht] hh",
fast 2 $ sound "bd sn [cp ht] hh"
]
swingBy x n divides a cycle into n slices and delays the notes in the
second half of each slice by x fraction of a slice. So if x is 0 it does
nothing, 0.5 delays for half the note duration, and 1 will wrap around to
doing nothing again. The end result is a shuffle or swing-like rhythm. For
example, the following will delay every other "hh" 1/3 of the way to the
next "hh":
A more literal weaving than the weave function. Given tabby threads p1 p,
parameters representing the threads per cycle and the patterns to weave, and
this function will weave them together using a plain (aka ’tabby’) weave,
with a simple over/under structure
Like contrast, but one function is given, and applied to events with
controls which don't match. unfix is fix but only applies when the
testing pattern is not a match.
The ur function is designed for longer form composition, by allowing you to
create ‘patterns of patterns’ in a repeating loop. It takes four parameters:
how long the loop will take, a pattern giving the structure of the composition,
a lookup table for named patterns to feed into that structure, and a second
lookup table for named transformations/effects.
The ur- prefix comes from German and
means proto- or original. For a mnemonic device, think of this function as
assembling a set of original patterns (ur-patterns) into a larger, newer whole.
Lets say you had three patterns (called a, b and c), and that you wanted
to play them four cycles each, over twelve cycles in total. Here is one way to
do it:
let pats =
[ ( "a", stack [ n "c4 c5 g4 f4 f5 g4 e5 g4" # s "superpiano" # gain "0.7"
, n "[c3,g4,c4]" # s "superpiano"# gain "0.7"
]
)
, ( "b", stack [ n "d4 c5 g4 f4 f5 g4 e5 g4" # s "superpiano" # gain "0.7"
, n "[d3,a4,d4]" # s "superpiano"# gain "0.7"
]
)
, ( "c", stack [ n "f4 c5 g4 f4 f5 g4 e5 g4" # s "superpiano" # gain "0.7"
, n "[f4,c5,f4]" # s "superpiano"# gain "0.7"
]
)
]
in
d1 $ ur 12 "a b c" pats []
In the above, the fourth parameter is given as an empty list, but that is where
you can put another lookup table, of functions rather than patterns this time.
For example:
let
pats = ...
fx = [ ("reverb", ( # (room 0.8 # sz 0.99 # orbit 1)))
, ("faster", fast 2)
]
in
d1 $ ur 12 "a b:reverb c:faster" pats fx
In this example, b has the function applied that’s named as reverb, while c
is made to go faster. It’s also possible to schedule multiple patterns at once,
like in the following:
let pats = [ ("drums", s "drum cp*2")
, ("melody", s "arpy:2 arpy:3 arpy:5")
, ("craziness", s "cp:4*8" # speed ( sine + 0.5 ))
]
fx = [("higher", ( # speed 2))]
in
d1 $ ur 8 "[drums, melody] [drums,craziness,melody] melody:higher" pats fx
Given a pattern of probabilities and a list of (value, weight) pairs,
wchooseBy creates a Pattern value by choosing values based on those
probabilities and weighted appropriately by the weights in the list of pairs.
wedge t p p' combines patterns p and p' by squashing the
p into the portion of each cycle given by t, and p' into the
remainer of each cycle.
> d1 $ wedge (1/4) (sound "bd*2 arpy*3 cp sn*2") (sound "odx [feel future]*2 hh hh")
whenmod has a similar form and behavior to every, but requires an
additional number. It applies the function to the pattern when the
remainder of the current loop number divided by the first parameter
is greater or equal than the second parameter.
For example, the following makes every other block of four loops twice
as dense:
A binary pattern is used to conditionally apply a function to a
source pattern. The function is applied when a True value is
active, and the pattern is let through unchanged when a False
value is active. No events are let through where no binary values
are active.
Use within to apply a function to only a part of a pattern. It takes two
arguments: a start time and an end time, specified as floats between 0 and 1,
which are applied to the relevant pattern. Note that the second argument must be
greater than the first for the function to have any effect.
For example, to apply fast 2 to only the first half of a pattern:
d1 $ within (0, 0.5) (fast 2) $ sound "bd*2 sn lt mt hh hh hh hh"
Or, to apply (# speed "0.5") to only the last quarter of a pattern:
d1 $ within (0.75, 1) (# speed "0.5") $ sound "bd*2 sn lt mt hh hh hh hh"
For many cases, within' will function exactly as within.
The difference between the two occurs when applying functions that change the timing of notes such as fast or <~.
within first applies the function to all notes in the cycle, then keeps the results in the specified interval, and then combines it with the old cycle (an "apply split combine" paradigm).
within' first keeps notes in the specified interval, then applies the function to these notes, and then combines it with the old cycle (a "split apply combine" paradigm).
For example, whereas using the standard version of within
In the following, the first pattern is the most likely and will play about half the time, and the last pattern is the less likely, with only a 10% probability.
An implementation of the well-known xorshift random number generator.
Given a seed number, generates a reasonably random number out of it.
This is an efficient algorithm suitable for use in tight loops and used
to implement the below functions, which are used to implement rand.
See George Marsaglia (2003). "Xorshift RNGs",
in Journal of Statistical Software, pages 8–14.
Rational numbers, with numerator and denominator of some Integral type.
Note that Ratio's instances inherit the deficiencies from the type
parameter's. For example, Ratio Natural's Num instance has similar
problems to Numeric.Natural.Natural's.