The Three Sirens
unsupervised selflearning robotic rock band
The story of the band begins with Aglaopheme, the slide guitar robot. I designed, built, programmed and testet the system in 1992. And successfully ran Aglaopheme for the first time during three days of "Electronic Arts Syndrom", Berlin in november that same year. Training of her neural networks began right there when the doors of that show opened. All artificial neurons were randomly initialized and the first visitors experienced a wildly out of control robot.
|The guitar robot turned out to be so exciting and so promising that I set out to build a three piece band around it to form The Three Sirens.
The newly created bandmembers were Peisinoe (single string bass) and Thelxiepeia (roto drum).
The Three Sirens premierd with a 3 day performance during "Mediale", Hamburg 1993.
The bass robot system operates in a similar fashion as the slide guitarnsystem. Only simpler because the bass has only one string.
The drum robot however only follows pre programmed patterns at system heartbeat. At an early development stage it became aparent that my neural network design can not control temporal patterns as it would be required for playing the drums.
I then added a voice to the lineup. ASCIIlizer, a set of neural networks, similar to those used by guitar and bass, that controls a bank of sampled phonems. The phonem were edited from recordings of my voice. Through endless hours of training the networks learned to control playback of these vocal fragments in volume, pitch and speed.
More rhythmic devices were added to the band. LynxArm for example is a small robot arm carrying a camera and bashing it rhythmicly into a block of wood. Camera images were used in a live video mix and the block carried a contact microphone.
Later I also added: kick drum, tomtom, hi hat, cymbal and shaker to the percussion section.
On several occasions I gave control over a standard synthesizer to Kohonen Networks. Thereby creating some more virtual band members to jam along.
The Three Sirens are not conventional musical instruments and they are not puppets but autonomous, self learning robot musicians. They are not influenced by human knowledge but explore their acoustic world and search for structure and corelation. Artificial neural networks control every aspect of the robot's activities and they only know what they hear. This means that the authorship for their music does not belong to any living being (no predefined material or musical knowledge is available to the system), all musical material is generated by self organizing learning processes in an improvising manner.
Aglaopheme, the guitar player
Aglaopheme's body is both, a robot with framework, actuators and
position sensors and a musical instrument (slide guitar) with six
strings, machine heads and a guitar pickup.
Six solenoids are used to either dampen and release or to pluck the strings.
A short piece of cable tie serves as a very durable plectrum that is moved across the strings by a linear motor to pick selected strings. The entire plectrum-drive assembly can be pulled away from the strings by a solenoid to allow the plectrum to drive to the desired strings without hitting any others.
A dc motor drives a sledge up and down the guitar neck. This sledge carries the slide and a small solenoid that pushes the slide against the strings if necessary. The sledge also carries a small video camera.
Two more video cameras are attached to the machine. One is located at the lower end of the guitar neck. The other one is mounted on top of a leg-solenoid.
Two big solenoids serve as legs to the robot. They can push up the entire framework in two different directions and thereby make the robot rock. This rocking motion also moves the whaa-whaa pedal up and down.
The other Members of the Band
Peisinoe, the bass player is a system very similar to the slide
guitar robot. It is only much more simple in that it only has one string
and therefore only one neural network to control it.
In order to vary the pitch, the robot uses one strong solenoid to adjust the tension of the string. The machine has two mechanisms to cause vibration of the string: a motorized bow and a slapping device.
Aciilyzer, the voice of The Three Sirens is still somewhat bodyless. Again, there is a neural network that takes audio spectrum data as sensorial input and that is in control of an sound generating device. But in case of this band-member the visual appearance of the machine is very boring.
The sound generating device is a database filled with vocal sound
samples of phonemes. Recordings of my own voice were used to set up the
phoneme collection. The neural network is trained by allowing it to
explore the outputs of the database. In a later state of learning the
vocal sounds fed into the network are mixed with, or even replaced by,
audio signals from the other members of the band. Intermixing the sounds
of the different instruments makes the individual systems responsive to
each other. In other words: they hear each other.
The percussionists in the band are operated differently from the
melodically instruments. The rhythm machines are basically just hooked
up to the system update frequency. If for example the band plays at 100
bpm, and the guitar- and the bass-system perform a learning cycle every
quarter note (roughly 6 analysis per second), the bass drum would make a
kick once per measure. All percussion devices can be programmed to act
with different repetition rates within this basic quarter note
Thelxiepeia, the drum was build around an existing rototom, a drum that was fashionable in the seventies. On this drum the player could change the pitch during playing by turning it. I motorised the rotational activity and added a mechanic drumstick.
LynxArm is a small five axis robot arm kit that I bought during a robotic exhibition in Vienna. After assembling the kit I changed one mechanic function and attached a small camera to what originally was the gripper. I also attached some sound pickups to the framework. The robot now knows two modes of operation. It either follows directly the spectrum analysis of the guitar sounds or knocks the camera against a small wooden board at system update tempo divided by four.
Audio recordings of The Three Sirens
|title||duration||low quality||high quality|
|extra laid back||2:07||0,99Mb||1,94Mb|
|robot rocks (extended version)||5:28||2,50Mb||5,01Mb|
|aglaopheme's solo 1||2:19||1,06Mb||2,12Mb|
|aglaopheme's solo 2||1:13||0,50Mb||1,12Mb|
|click filesize for playing|