breakthru

i am writing this post to set down that a breakthrough has occurred. i have written a patch that achieves a level of machine listening and autonomy of note event selection hitherto unattained in any of my previous work with computer music. this is the culmination of 2 years of struggling with conceptual limitations and hacks.

allow me to explain (or try minimally and fail). sorta how tubby now knows how to speak, i have given similar autonomy to shrod. a large part of it is his improved filterbanks. now he's got 539 analysis bands, localized in time and frequency. he also has 3 stages of amplitude resolution, and the microsounds he triggers can be scaled accordingly or be completely different depending on these stages. I guess that's what I mean in terms of autonomy. The data coming in to this puppy gets categorized and analyzed in ways that have come out of the study of psychoacoustics. for example his analysis bands are logarithmically spaced (at least for now) at 50 bands/octave. what shrod does with these osc messages is entirely up to me as well, so i can do realtime transforms which range from attempts at reconstructing the data (a venture whose success hinges upon what the material is and how it relates to the reconstructing 'vocabulary') to pattern detection and counterpattern deployment (realtime software automata that can realize any number of strategies such as tone clusters, counterpoint, rhythmic motifs, etc). this thing has a view from within the sound and is able to not just change the stimulus but produce strategic responses. Eventually, he will use a bank not of constant-q filters, but of cq driven phase vocoders arranged to do multiresolution analysis across multiple time/frequency scales. although at that point each octave will be divided linearly, over each octave the stimulus will be analysed with an appropriate window and provide identically scaled time/frequency resolution. The phase vocoders will similarly not be driving cosine summation oscillators but triggering note events, so the strategy will be similarly variable. This will also give us data about the phase structure of the stimulus, making wavelet and matching pursuit operations extremely straightforward. at this moment i have attempted a few tricks with him, including reproducing him with a live updating wavelet shape derived from an envelope function and the current stimulus, transposed across the spectrum. i have also tested out assigning different wavelet shapes to each of the 3 levels of dynamic resolution. also i have used a few synthetic wavelet shapes such as banded whitenoise with a formant (always a favorite) and straight sinusoid tone pips with a von hann envelope.

I hope to use this patch as an automaton to trio with tubby and me, since tubby i see as his hardware hacker, physical computing, evil jamaican twin.

I am also experimenting with the idea of remixing (or something) this album I recorded in 2004 on a beautiful, impeccably mic'ed concert grand. It was called "Animals Should Wear Clothes", and it is largely solo piano with some hand percussion and bowing the strings, etc. I have already produced a few starting points, using shrod. there's actually a piece on the album called "Schroedinger's Cat" which I have started working with. It was unconscious I swear! One of these remixes was my CommLab final because I was so excited about this thing for the past few waking cycles. I'll post links to the music when I upload it in some fashion or another.

in other news, i've lost my mind.

Leave a Reply

Your email address will not be published. Required fields are marked *