Skip to content

Changelog

Release Notes

2.0.0 - Beta 4

Breaking Changes

Flip-flopped back to using strings for name lookups when calling from python. (Internally flags are still used in many places for performance.)

So use eg dsp.wt('sine') instead of dsp.wt(dsp.SINE) or sound.pan(0.1, method='gogins') etc.

Wavetable / window options are:

  • sine
  • sinein
  • sineout
  • cos
  • tri
  • saw
  • phasor
  • rsaw
  • hann
  • hamm
  • black or blackman
  • bart or bartlett
  • kaiser
  • rnd
  • line
  • hannin
  • hannout
  • gauss
  • gaussin
  • gaussout
  • pluckin
  • pluckout

  • square

  • sinc

Panning types:

  • constant
  • linear
  • sine
  • gogins

Features

  • Removed all the old examples and wrote the first tutorial script.
  • The first set of soundpipe modules are now available via the fx module!
    • fx.lpf, fx.hpf, fx.bpf, and fx.brf butterworth filters.
    • fx.compressor... a compressor.
    • fx.mincer a phase vocoder with independent control over pitch and speed.
    • fx.paulstretch... paulstretch.
    • fx.saturator a saturation distortion.
  • A new Waveset datatype for Trevor Wishart-style microsound synthesis and easy creation of wavetable stacks for 2d oscs. Featuring:
    • Waveset substitution
      • Single waveform substitution (replace all wavesets with a sinewave, for example)
      • Waveset collection substitution (replace wavesets with a collection of wavetables or another waveset)
      • Waveform morph substitution (replace wavesets with an lfo-controlled morph through a wavetable stack) (in progress)
    • Waveset normalization
    • Waveset time-stretching
    • Waveset reversal
    • Waveset retrograde (reversing the contents of the wavesets while preserving their order)
    • Waveset harmonic distortion (in progress)
    • Waveset transposition (in progress)
    • Waveset morphing (in progress)
    • Waveset inversion (in progress)
  • Broke oscs up into smaller units and added a few new ones consisting of:
    • Osc which is a simple wavetable osc
    • Osc2d a simple 2d wavetable osc
    • Pulsar a pulsar synthesis implementation (now with pulsewidth modulation!)
    • Pulsar2d the same which accepts optional wavetable stacks for the wavetable and window params
    • Fold an implementation of an infinite wavetable folder
    • Pluck a basic implementation of a plucked string physical model which can be fed with an arbitrary wavetable impulse
    • DSS an implementation of dynamic stochastic synthesis (in progress)
    • Alias a single-sample aliasing pulsetrain osc
  • Total rewrite of grains.GrainCloud, now grains.Cloud (and SoundBuffer.cloud)
    • Uses mincer for pitch shifting
    • Grainlength and grain density are no longer tightly coupled: no more density param. It is replaced with a grainlength value/wavetable and a grid value/wavetable.
    • Grain masking
  • Many params can be given as either a fixed-value float (like 1.3), a built-in wavetable flag (like hannout or rnd), or a wavetable-like object (a list, numpy array, sound buffer, etc). EG a cloud frozen in time Cloud(position=0.75), advancing linearly through time Cloud(position='phasor'), or interpolated through four points in time Cloud(position=[0, 1, 0.25, 0.75]).
  • A simple envelope follower interface on SoundBuffer which produces a Wavetable via SoundBuffer.toenv(), also available as fx.envelope_follower()
  • A SoundBuffer-to-Wavetable shortcut via SoundBuffer.towavetable()

Bugfixes

  • Some cython extension packaging improvements
  • Many misc bugfixes, I lost track... probably also at least a few shiny new bugs as well

2.0.0 - Beta 3

Features

  • Basic FIR filter with filters.fir or fx.convolve
  • SoundBuffer normalization with fx.norm
  • Get SoundBuffer magnitude with dsp.mag
  • Single and multitap delays with fx.delay and fx.mdelay
  • Some new built-in wavetable.window types: dsp.SINEIN / dsp.SINEOUT, dsp.HANNIN / dsp.HANNOUT for fades
  • More flexible frequency table creation from arbitrary scales, tunings and scale bitmasks with tune.tofreqs

Bugfixes

  • Fix phase overflow in interpolation._linear_point
  • Fixed a nasty bug when loading mono soundfiles from disk.

2.0.0 - Beta 2

Features

  • Point interpolation with interpolation.linear_point

Bugfixes

  • Examples can be run from anywhere
  • Interpolation fixes

Performance Optimizations

  • Faster ADSR wavetable generation
  • Faster pitch shifting
  • Faster interpolation
  • Faster grain cloud generation
  • Some misc SoundBuffer performance improvements (more to come)

2.0.0 - Beta 1

Features

  • Added fx module
  • Added first pass fx.go granular overdrive effect. See examples/fxgo_example.py for usage.

Bugfixes

  • Fixed a packaging issue preventing the tune module from loading.
  • Better overflow handling in SoundBuffer.adsr and wavetables.adsr
  • Fixed a bug with SoundBuffer.remix when mixing to a single channel
  • Fixed a bug during Wavetable initialization when using wavetable flags to create a window.

2.0.0 - Alpha 9

Final feature releases / big api changes before going to beta.

  • Breaking changes:
    • Durations for most APIs are now given in seconds (floats) rather than integer frames. len(SoundBuffer) still returns a length in frames per python convention, and slicing into a SoundBuffer is also still done by frame (and channel) but there is a new SoundBuffer.dur property to get duration in seconds as well.
    • Wavetables are no longer specified with string names, instead built-in flags which are available on both the wavetable and dsp modules are used. Eg to apply a sinewave envelop: sound.env(dsp.SINE) instead of sound.env('sine'). The wavetypes available are SINE, COS, TRI, SAW (which is also aliased to PHASOR), RSAW (reverse sawtooth), HANN, HAMM, BLACK or BLACKMAN, BART or BARTLETT, KAISER, SQUARE, and the RND flag to select one at random.
  • Osc changes:
    • Added 2d wavetable synthesis (similar to max/msp 2d.wave~) to Osc plus example script
    • To create a 2d Osc, use the stack keyword arg on initialization: Osc(stack=[dsp.RND, [0,1], dsp.SINE], lfo=dsp.SINE)
    • Osc wavetables may be:
      • an int flag for standard wavetables (dsp.SINE, dsp.TRI, etc)
      • a python list of floats ([0,1,0.5,0.3])
      • a wavetable (wavetables.Wavetable([0,1,0,1]))
      • a soundbuffer (soundbuffer.SoundBuffer(filename='something.flac'))
    • 2d wavetable stacks are a python list of any combination of the above.
    • The same types are acceptable for:
      • wavetable (the basic waveform)
      • window (an optional window to apply to the waveform wavetable - useful for eg pulsar synthesis)
      • mod (the frequency modulation wavetable)
      • and lfo (the 2d modulation wavetable)
  • SoundBuffer changes:
    • Added remix for remixing a soundbuffer from N channels to N channels.
    • Panning algorithms operate on arbitrary numbers of channels (but use same algorithms applied to odd & even numbered channels instead of left & right)
    • Return a reversed copy of a soundbuffer with sound.reversed() or reverse in place with sound.reverse()
    • New ADSR envelopes with sound.adsr(a=1, d=1, s=0.5, r=1)
    • Generate a GrainCloud from a SoundBuffer with sound.cloud()
    • Clip samples to min/max with sound.clip(minval=-1, maxval=1)
    • Taper ends of sounds (linear fade-in, fade-out) with sound.taper(length)
  • ADSR wavetable generator with wavetables.adsr(a=100, d=100, s=0.5, r=100, 1024)
  • New Wavetable type for SoundBuffer-like operator-overloaded wavetable manipulation & composition
  • New GrainCloud wavetable-driven granulator. See the examples/swarmy_graincloud.py example for more.
  • GrainCloud-driven pitch shift without time change (sound.transpose(speed)) and time stretch without pitch shift (sound.stretch(length)) methods for SoundBuffer.
  • dsp.cloud(SoundBuffer, *args, **kwargs) shortcut for GrainCloud creation.
  • Read wavetables from 1 channel sound files with wavetables.fromfile
  • Added a simple helper for async rendering with multiprocessing.Pool
  • SoundBuffers can now be pickled (enables passing them between processes)
  • SoundBuffer can be initialized (and spread across channels) from a normal python list

2.0.0 - Alpha 6-8

This was meant to be a feature-only release, to add the final round of features before going into beta / bugfix mode. Instead I switched from using numpy arrays directly to a first pass of a more general typed memoryview approach, and moved some more things into cython.

Also, hoo boy was pip install pippi ever broken. It should be working now.

2.0.0 - Alpha 5

Bugfix release. Fixed an idiotic regression in SoundBuffer.

2.0.0 - Alpha 4

New in alpha 4:

  • More speed improvements!
  • Linear interpolation option for pitch shifting
  • Pulsar synthesis with Osc
  • Support for importing Scala .scl tuning files (Mapping file support coming later...)
  • Waveform visualization with graph.waveform
  • SoundBuffer.fill returns a copy of the sound instead of altering it in place
  • Sampler abstraction for Osc-like treatment of samples and banks of samples

2.0.0 - Alpha 3

Optimizations and improvements in this release:

  • Much better performance for wavetable and granular synthesis
  • Improvements and additions to the rhythm modules
    • Better handling of rhythm.curve which now takes a length param instead of an obscure combination of multipliers
    • rhythm.curve can now be provided a custom wavetable
    • Added MPC swing helper for onset lists (via rhythm.swing)
    • Added a euclidean rhythm generator rhythm.eu
    • Added a pattern generation helper rhythm.pattern
    • Added pattern-to-onset and string-to-pattern helpers
      • Patterns are the same as pippi 1 (I may even just port some code) and can be in a few forms:
        • String literals with ascii notation eg: 'xx x- x' which is the same as 'xx.x-.x'
        • Lists of 'truthy' and 'falsey' values eg: ['1', True, 0, False] which is the same as 'xx..'
  • Misc bugfixes:
    • Fix random param for wavetable.window and wavetable.window
    • Fix bad params for wavetable.window and wavetable.window -- falls back to sine in both cases

2.0.0 - Alpha 2

This release includes a few missing pieces to core functionality including:

  • A crude squarewave wavetype for the wavetable osc!
  • Custom wavetables for the wavetable osc and window/wavetable generators! See the simple_custom_wavetable.py example for use with the wavetable osc.
  • A simple non-interpolating speed method on SoundBuffer for pitch shifting sounds
  • A set of rhythm helpers in the rhythm module useful for constructing onset / timing lists.
    • rhythm.curve which lets you map any of the window types to a list of onsets -- check out the simple_snare_bounce.py example in the examples directory.
  • Some more example scripts including:
    • simple_snare_bounce.py Demoing the rhythm.curve helper
    • multi_snare_bounce.py A more interesting variation on the snare bounce example
    • simple_custom_wavetable.py Showing a user-defined wavetable used with Osc
    • synth_chords.py Using the tune module with Osc to create a simple chord progression

2.0.0 - Alpha 1

This is the initial alpha release of pippi 2 -- which is very barebones at the moment, but already pretty functional!

Beware: the behavior of core functionality and features will probably change throughout the alpha releases of pippi 2. I'll try to document it here in the release notes.

This release provides:

  • SoundBuffer abstraction for reading/writing soundfiles and doing basic operations on sounds.
  • Osc abstraction for simple wavetable synthesis.
  • Initial set of built-in wavetables for windowing (sine, triangle, saw, inverse saw) and synthesis (sine, cosine, triangle, saw inverse saw)
  • Set of panning algorithms and other built-in sound operations like addition, subtraction, multiplication, mixing (and operater-overloaded mixing via sound &= sound), dubbing, concatenation.
  • A small set of helpers and shortcuts via the dsp module for loading, mixing, and concatenating (via dsp.join) sounds.
  • Basic granular synthesis and wavetable synthesis examples.