Computers Learn Sound: Creating Timbre Through Neural Networks

This is on going research in using neural networks to generate variations of natural sounding timbre.  For this research the Supercollider programming environment is used to train a neural network for one type of timbre.  After the error amount is small, approximately 0.05 the neural network is told to generate a sound file based on it’s learning.  In theory the sound file should sound related to the learning data i.e. learning data of snare drums should cause the network to produce a snare drum sound.  Then new sounds that are not part of the learning set are slowly introduced into the system.  At this point the network should start to think that the non-related sound is related.  When the neural network is then asked to generate a sound file, the sound file takes on new characteristics based on the sounds that are added to the learning process.

Chemotactic Response Sonification

By Justin Peruzzi, Maxwell Tfirn and Dr. Rosanne Ford

Chemotaxis is the study of a cell or organisms motion in response to a particular substance.  For this research, video data is recorded from a microscope without a concentration of a substance.  The video is then played back through custom software that reads each video frame as a sonographic image.  This in turn creates sound that can be analyzed for spectral flux, noise, flatness and more.  A substance is then added to the slide, video recorded and ran through the same process.

If the cells are responding to the substance the density of the sound will change.  In our results the control will create random dense sounds and have a high spectral flux and noise.  When an attractant is added that slows the motion of the bacteria down, the sound will be less dense, and the spectral flux and noise will be lower.  This shows that audio can be used to understand the response of cells that may not be visible to the naked eye.

Abstract can be found here

Image to Sound Mapping

For this research a suite of max/MSP patches and abstractions was created that allows the user to draw a sonogram, erase a sonogram, import a picture as a sonogram, spectrally analyze sound and time stretch or freeze an instance of the spectrum.  This research uses all native max/MSP objects and allows the user to delve into spectral music in a interdisciplinary way by adding a graphic design element to music composition.

SpectrumDraw is a program that anyone can draw in and produce a sound based on the drawing.  It allows the user to draw freely or trace a picture and hear what it sounds like.  This program is inspired by Xenakis’s UPIC system and MetaSynth.

Note: Gatekeeper on Apple computers needs to be turned to “Anywhere” for software to run the first time it is opened.

Download SpectrumDraw Here

Erase Spectrum is a program that allows the user to draw onto a white screen.  The drawing will cut holes in the sound which is comprised of a quasi aperiodic (White Noise) sound.  If the sound is recorded and graphed back the image will reappear.  This program can be used to add messages or pictures into a sound.

Note: Gatekeeper on Apple computers needs to be turned to “Anywhere” for software to run the first time it is opened.

Download Erase Spectrum Here

SonoPlay is a max/MSP abstraction that can manipulate a spectrum and analyze different aspects of the sound.

Download SonoPlay Here

Sound to Score

This project originally acted as an extension of the “Image to Sound Mapping” research and turned into its own musical research.  This project originally was a way to take the sound produced by the “Image to Sound” research and notate it to understand the data in musical terms.  After this was achieved, the next step was to research how spatial notation can be interpreted into traditional music notation.  Finally a way to play back a sound file and notate it in real-time was created.  This research is a compositional aid for notational possibilities that accurately represent time.

Three programs were created using Max/MSP and the Bach external library to realize this.  The first program, Image Score, loads an SDIF file and the user can see how what comprises the spectrum is in terms of pitch and changes over time.  The program can also show microtones up to an 8th tone.  The second piece of software, SpaSco Notation, allows the user to spatially notate any score or SDIF in any key and clef and quantize it to any rhythmical value such as 1/4 or 1/26.  The program then gives the notation in a traditional score format that can be saved as an XML file and imported into a program such as Finale.  Finally, the last piece of software, SoundScore, allows a sound file to be loaded and analyzed for peak frequencies to be notated on a music staff.  SoundScore can also be used in real-time to capture and notate sound.  The file can then be exported as an XML file for Finale to use.

Note: Gatekeeper on Apple computers needs to be turned to “Anywhere” for software to run the first time it is opened.

Download Image Score Here

Download SpaSco Notation Here

Download SoundScore Here

Clave Pattern

For this research I developed a Max/MSP program for the Kubovy Perception Lab at UVA to test rhythmic identification in spoken sentences. The software plays words and rests in patterns, and the task is for the subject to determine the pattern’s starting point. The software also saves the answer to all trials.

Sample and Hold Granularizer

The Sample and Hold Granularizer or S.A.H.G. is an instrument that allows the user to granulate a sample and control the grain length, inter-grain time, variation of time, envelope, panning, LFO and filters.  S.A.H.G. was part of a project to create various controllable sound makers for a laptop ensemble composition. The main focus of this research was to be able to control the instrument without touching the computer i.e. use a controller to manipulate every aspect of the instrument.  This instrument is intended to create controllable, singular granular lines in laptop ensemble settings.  Built into this instrument is a midi matrix that can be used to connect any midi controller and assign various different knobs and sliders to the instruments parameters.

Note: Gatekeeper on Apple computers needs to be turned to “Anywhere” for software to run the first time it is opened.

Download S.A.H.G. Here

Real-Time Sequence Maker

RhythmMachine is a digital instrument that allows the user to control a bass, snare and two noise sounds in a sequence fashion.  Each sequence has its own tempo control, filtering and envelope that can be manipulated in real-time via a midi controller. This software instrument was originally designed to study non-quantized rhythm sequences and real-time control of rhythm for performance.  However, after it was developed it turned into a laptop instrument that can create intricate rhythmical patterns for live performances.

Note: Gatekeeper on Apple computers needs to be turned to “Anywhere” for software to run the first time it is opened.

Download RhythmMachine Here


Stretcher is a tool for taking a sound and repeating a part of it indefinitely.  This research was focused on smooth time stretching of a sound that also allowed the user to select part of the sound sustain  This patch was also created to create spectral sonorities between computer sounds and instrumental sounds that would normally be on the micro level.  Being able to play back or stretch parts of a sound that happen in a fraction of a second enables subtle detail to be brought to focus and used in the composition.  This tool uses a graphic sonogram and allows the user to select an individual frame for playback.  The user can also control the time it takes to play all of the frames as well as randomize the frames.

Downloaded Stretcher Here