Kyma [Scaletti, 1991,Scaletti and Johnson, 1988,Scaletti, 2002] is a visual sound design language that was started by Carla Scaletti at the Illinois University at Urbana Champaign and is now commercialized by the authors through Symbolic Sound Corporation [www-SymbolicSound, ]. Although Kyma is much more than a graphical application and it contains a complete conceptual model very much in the line of the already commented MODE or Siren, it is best known and only used as a graphical application, and for that reason it has been included in this category, even at the risk of disappointing the author who clearly states that Kyma would still be Kyma without the graphical interface[Scaletti, 2002] but also qualifies it as a ``visual language'' [www-SymbolicSound, ].

Kyma is an object-oriented music composition and sound synthesis environment written in Smalltalk-80 that can be used with a microprogrammable digital signal processor called Capybara. According to the authors [www-SymbolicSound, ] Kyma is being used to do sound design for music, film, advertising, television, virtual environments, speech and hearing research, computer games, and other virtual environments. Although this may be so, the truth is that Kyma has been designed and it is mostly used as a music composition tool.

The first version of Kyma was started in 1986 on a Macintosh. In 1987 it was modified to make use of the Platypus signal processor. In 1989 they decided it made no sense to keep the project in the university so they founded a company, Symbolic Sound Corporation, that started operating on their student apartment and moved to their first office in 1992. Since the first version of Kyma in 1986, there have been five major software revisions, a port from MacOS to Windows in 1992, and ports to five different hardware accelerators including the currently supported Capybara. Currently Kyma runs either on a Macintosh or PC and on the Capybara, a general-purpose multiprocessor computer that can support from 4 to 28 parallel processors.

Kyma provides ways to create and manipulate sound objects graphically with real-time sonic feedback via software synthesis. In order to offer real-time synthesis without having to compromise flexibility, the authors decided for a specialized hardware DSP for being both efficient and programmable. But, as the authors acknowledge, the main reason for using the Platypus platform was that in the computer with Smalltalk they couldn't even get 20k samples per second.

When designing Kyma, the authors intentionally kept away from Music-N languages and music notation based systems [Scaletti and Johnson, 1988]. According to them the main reasons that on the one hand, Music-N languages seldom offer immediate feedback and can be frustrating for composers trying to experiment new sounds and they make it difficult to control higher level-aspects of music such as phrases or lower-level features such as those related to timbre, which have to be controlled from a different file. And on the other hand, environments based on music notation do not offer enough flexibility as an acoustic event cannot be fully specified with traditional music notation.

In Kyma there is no clear distinction between Instrument and Score. Everything in Kyma, from a single timbre to the structure of the whole composition is a Sound Object. The sound object in Kyma is inspired on the objet sonore of Pierre Schaeffer [Schaeffer, 1966]. A Sound Object can be manipulated, transformed, and combined into new Sound Objects. Objects that were encapsulated into another object can be brought back to the top level object, and top level objects can be combined and hidden in a yet higher-level object. Sound objects are uniform so any given Sound object can be substituted for any other Sound object.

A Kyma Sound Object can be either a SoundAtom or a Transform. While a SoundAtom is a regular sound or collection of samples, a Transform is the sound result of applying a given function to its subsounds. In this sense, a Sound in Kyma is represented as a directed acyclic graph (DAG) with a single root node. Each edge in the graph represents the relation ``is a function of''. A subsound can be shared among several superSounds. A Sound DAG is similar to an expression tree in that the evaluation of the higher nodes depends on the result of the lower nodes.

In order to hear a Sound object it is necessary to ``evaluate'' it, that is, convert it to a sample stream. Every sound object knows how to compute its next sample. The nodes of the DAG are evaluated in post-order. When a Sound object DAG includes Delay nodes the DAG is first expanded into a series of time-tagged DAG's.

A SoundAtom has no subSounds. For instance a LiveSound is defined as the input from the analog-to-digital converter. A PotentialSound is one that does not respond directly to a Play message. When it receives such a message it creates an new Sound from its Subsound and then sends it the play message. A PotentialSound node expands into a subgraph during evaluation. An interesting example of PotentialSound is the FrequencyTransform, it contains as an attribute a function of time, frequency and the duration of its subsound.

A Kyma Lifted object is a Sound object with variables. A lifted object can work as an ``instrument'' that is instantiated and scheduled from a score as in the Music-N style languages. In Kyma a MusicN is a Sound object whose parameters include a score and a collection of subsounds to be used as instruments. Any variable parameter of a subsound can be set from the score. In the score language an event is specified as the subsound name, a start time and any number of (parameter, value) pairs. Parameter values can also be subsounds.

A Kyma Transform can be unary or N-ary. An N-ary Transform has an OrderedCollection of sounds (named subSounds). The primary N-ary transform are Mixer, Concatenation and Multiplier. On the other hand, an unary transform has a single sound in its attribute subSounds (e.g. Delayed, Amplitude Scaled...).

It takes a finite amount of time to compute a sound object, therefore, real-time cannot be guaranteed. Kyma attacks this problem in two ways: one option is to lower the sample rate and try again, the other is to allow the output samples to be stored on disk (the sample file on the disk can be treated just like any other object).

The conceptual object-oriented model in Kyma is a central issue and it is in fact the only thing that has not changed since its first versions. According to the authors, this demonstrates its flexibility and how well it suits the creative process. For example, in the first versions although structures could be time-varying, parameter values were constant. But they realized that making parameters event-driven the language would be more interactive and would open to external control such as MIDI. In version 4.5 they added an event language. The improvement on the quality of the new sounds demonstrated it had been a good decision. But the original data structure was robust enough to accommodate to this major shift.

One of the ideas behind Kyma is that one can easily plug new algorithms. Kyma has proven open enough to accommodate new algorithms that the authors have developed during the years. It is important to note that Kyma is not designed to allow the user to extend the environment by adding new objects and does not offer access to its source code. This is one of the reasons we cannot classify as a framework.

But apart from its conceptual model, Kyma has an extensive and flexible user interface. The current graphical representation has evolved over the years. The abstract structure came first, and the graphics evolved in order to represent the structure.

In the first versions a Sound was specified as Smalltalk code. This quickly evolved into a ``selection-from-list'' interface. In order to reduce the need for typing, the interface was changed to the ``Russian Doll'' style where Sounds were represented as containers of other sounds. Double clicking opened a new window showing this sound. This interface was good at reveling the recursiveness but not the overall structure. Then it was changed to the box interconnection paradigm ala Max with flow direction from top to bottom, representing a DAG (directed acyclic graph). Later the direction was changed from left to right. Although the underlying structures did not change, changing something as simple as the direction made people understand the structure differently and therefore create differently. Apart from the central structural representation, Kyma also includes a timeline representation.