Evolving Sonic Environment
This is an ongoing collaborative work between Usman Haque, Haque Design + Research and myself.
In this version we have lowered the operating frequency to a more audible level and incorporated deeper modulation of tone in the sounds emitted by each device. The devices are now much more complex and are able to display a higher level of interactivity.
As with previous versions we use acoustic coupling between the devices, but with peak sensitivity around 3.7KHz, about the same peak response as the human ear. The sounds emitted are able to sweep across a wide range of frequencies to include 3.7KHz at roughly the centre. The devices are able to produce sounds with upward and downward intonation, this being controlled by a network of seven neurons per device. In its current form around 12 devices are hung, distributed in a room at different heights and attitudes and allowed to interact with each other and viewers entering the room. Each device is entirely constructed of analogue components and are continuous in the levels and timings of the signals passed within them, there is no program code, and no central processor in these devices. A functional description of a device is given below;
Each device has an 'ear' constructed of a condenser microphone housed in a metal tube cut to length such that it behaves as a quarter wave resonator around 3.7KHz. The signal from this microphone is then amplified and applied to a pair of independent tunable tone decoders. The outputs of these tone decoders become low where sounds within the (quite narrow) pass band of the tone decoder is detected. These signals are passed onto the neural network stage.
|Evolving Sonic Environment installed at Netherlands Media Art Institute, Amsterdam, Netherlands, 2nd June - 14th July, 2007|
A group of four analogue neurons decodes the signal into rising or falling tones. The first two neurons buffer and invert the signals from the two tone decoders. There are connections to two further neurons such that one of these neurons receive delayed signals from one of the two buffer neurons and signals directly from the other buffer neuron, this allows the neuron to detect the change over time of the signals from each of the tone decoders. If the two tone decoders are tuned such that they are close in frequency, but with only a small degree of overlap in their pass bands, the output of one of these neurons will become high when tone sweeps over time in a particular direction are applied. The outputs of these two neurons now code signals with either rising or falling intonation. The signals from these tone slope detection neurons are now applied to the inputs of two long and two short term analogue memory cells. The short term cell charges in a fraction of a second and discharges over around 10 seconds and acts to stretch the brief signals from the tone slope detection neurons. The long term memory cell has a time constant of around 60 hours and acts to store a a history of the signal dynamics detected by the device. Two further neurons now summate the signals from the memories and apply them to the output circuit to produce rising or falling tones. These two neurons also inhibit each other such that one neuron tends to become dominant at any one time. These neurons also have self inhibitory circuits that tend to suppress or down regulate their activity when they have been stimulated for an extended period. There is a final neuron that modulates the amplitude of the output signal and serves to silence the device when there has been excessive excitation from previous stages. The overall result of the excitatory and self regulating connections is to produce a system that seeks to achieve and equilibrium with the devices (or humans etc) that it interacts with.
Finally the output stage is a pair of oscillators, one tuned to around 3.7KHz and the other modulated at very low frequency (1-0.01Hz) by the outputs of the neural network circuit. The low frequency oscillator then modulates the pitch of the high frequency oscillator, the output of which is then amplified and applied to a small speaker.
Versions I & II
The aim behind this project was to create an adaptive system of sonically coupled devices able to interact with people and able to determine their presence or absence with no explicit sensors included in their design for this purpose. The devices communicated using very high pitched acoustic waves at around 16 KHz, just on the edge of most people’s audible range. This high frequency was employed so that, given the diameter of the transducers used (~60mm), the signals from each device would be beam like with a spread of around 20 degrees or so. This allowed complex reflections of the devices sounds around the room in which they were displayed, people entering the room would then disturb, block and reflect these sounds. The devices themselves were in a constant signaling each other with ‘chirp’ like noises - the rate of chirping indicating the internal state of a particular device. Each sonic device was completely autonomous in its function, requiring only a DC voltage supply to drive it. Each device consisted of an ‘ear’, a tone decoder, a neuron circuit and a ‘voice’ circuit, a little more detail on the function of these is given below;
The Ear circuit consisted of a microphone within a short horn like enclosure approximately tuned to around 16 KHz and a band pass filter/amplifier circuit, again tuned approximately to the 16 KHz region. The output of this circuit was connected to the tone decoder circuit.
The tone decoder circuit consisted of an oscillator tuned to around 16 KHz and phase detectors arranged so that if an in band signal was detected then an output voltage would be produced. This circuit was coupled to the neuron circuit as an excitatory input.
A spiking neuron realised in analogue electronics having an integrate and fire mechanism consisting of a capacitor and a voltage controlled switch such that when the capacitor became charged to a certain point it would be discharged almost completely with the result of outputting a voltage spike. The circuit also contained a memory element that modulated the firing threshold of the neuron over time such that the more often the neuron was triggered to fire that it became easier to fire, by this Hebbian learning was allowed to occur at the neuron circuit. If the circuit became too easy to fire then firing became inhibited - effectively the neuron became exhausted, or ‘bored’. This stopped the network activity from spiraling out of control. The neuron was also biased to fire at a slow rate even when not stimulated to do so from outside signals. The output of the neuron circuit was applied to the voice circuit.
|Evolving Sonic Environment installed at NTT InterCommunication Center [ICC], Tokyo, Japan,15th September - 26th November, 2006
Left: devices installed in the space. Right: Spectrum display of acoustic activity.
Voice circuit consisting of an oscillator running at around 16 KHz being frequency modulated by the voltage spikes from the neuron circuit. The output of the oscillator was then smoothed to remove excessive harmonic content, amplified and used to drive a small transducer.
The devices were arranged pseudo randomly in a room and connected to a voltage supply. Visitors were invited to view the devices and were free to move around them. A computer was set up to record the aggregate room sounds and to monitor for the presence of people for later offline analysis. It was found that the network fell into spontaneous patterns of resonance and that with the presence of people in the room these patterns were altered and destroyed, then when the room became vacant again new patterns of resonance emerged - similar but not identical to the original patterns. The signal passing between the sonic devices was seen to adapt over time, becoming faster to settle back into stable resonance.