Unit-generator graph as a generator of lazily evaluated audio-vector trees

Computer music systems normally implement a unit-generator graph as a network of unit generators, through which audio vectors are streamed and processed. This paper proposes an alternative implementation technique for unitgenerator- based sound synthesis, which views a unit-generator graph as a generator of audio-vector trees to be lazily evaluated. The simplest implementation of this technique allows to process sound synthesis and sound-control tasks in different threads even in a synchronous computer music system, making real-time sound synthesis more stable by amortizing the time costs for sound-control tasks over DSP cycles, while maintaining the low roundtrip latency between the audio input and processed output at the same time. We also extend the discussion to the possible extensions of our technique for parallelization, distribution, and speculation in real-time sound synthesis. The investigation into such a novel implementation technique would benefit further research on high-performance real-time sound synthesis.