I was interested in what sort of ouput we would get from training a wavenet encoder on many voices but treating them as a single voice.
Most usages segment this dataset by person, gender and country. I used the GMU global accents database for training.
I have been working on a project to use a sample dataset from the GMU Accent project
to create audio and music. The project has over 2600 samples exploring accents from across the world. Everyone is reading the same text which
highlights all the sounds of English. There are a huge variety of speakers and native languages with rich meta data for querying.
I've been exploring how this dataset could be used to explore communication and language, interacting with human dancers
controlling sample parameterisation through sensor feedback.
In exploring the data I did some simple DSP onset detection to extract all voices saying "train station".
While very simple, it is deeply beautiful hearing such a wide pallet of voices.
I've been exploring ways to deform meshes in 3D. The best way in the end was to use a terrain map to alter the mesh positions of a sphere. The terrain map was a blackwhite png where darkness maps to height.
Effectively rendering a mountain range into a 3d mesh.
Controlled by Emacs and SonicPi. Also using a focus blur for added effect.
An attempt to create an eye like shape with the edges/tentacles feeling like origami paper folding. Achieved through increasing light intensity, a cartoon edge like shader applied as a postFX and repeatedly resting the laws of physics with the edges/tentacles simulation.
For some unknown reason I'm slightly obsessed with Soprano samples. I've been playing around with the idea of the human ear seeking meaning in a repeated phrase that still has an element of a human voice.
I'm using wavetable playback manipulation, playhead speed and some added FM synthesis on a single 4 second Soprano sample.
I've been thinking on some advice given by Deru. "Turn organic sounds into electronic and electronic into organic" Based on this I've been creating an instrument from two classical instruments sounds, a single 1 second sample of a Soprano singing a latin phrase and a single 1 second sample of a Violin. Added together with a lot of granular synthesis. Part of a new piece of music I'm working on.
I recently added into ShaderView (https://github.com/josephwilk/shaderview) Post processing filters. Effects like bloom, fisheye, mirroring, noise and grain effects. Two of my favourite examples generated live.
Method missing, the method that does not exist is a magical and confusing concept. I wanted to draw attention to how counterintuitive this method is and the pain of many programmers trying to understand the behaviour of code using it.
When searching text in Emacs, matching terms are highlighted. I wanted to explore creating visual patterns through writing regular expressions.
I compressed a section of performance code and then in emacs created different searches.
I wanted to expand the expressiveness of patterns with Emojis rather than text. To add an extra level of meaning
for the audience and for my own mnemonics of mapping different types of percussive hits. Parameters like attack, release, velocity and sustain being different Emojis.
Converting live metric data from SoundClouds production systems into a mesh in OpenFrameworks. The connection of mesh vertexes is slightly random which creates interesting shapes between graphs peaks and troughs.
In most performances when I live code visuals and music the controlling code is overlaid over the top of the visuals. I've been experimenting with how to bring the text into the visuals rather than just sitting on top of it.
To do this I used OpenFrameworks I created meshes from black and white screen shots of text. Then adding noise to the text mesh creating interesting distortions. Making the text feel more organic and alive while still being readable.