Engineering Variance: Software Techniques for Scalable, Customizable, and Reusable Multimodal Processing

This article describes four software techniques to enhance the overall quality of multimodal processing software and to include concurrency and variance due to individual characteristics and cultural context. First, the processing steps are decentralized and distributed using the actor model. Second, functor objects decouple domainand applicationspecific operations from universal processing methods. Third, domain specific languages are provided inside of specialized feature processing units to define necessary algorithms in a human-readable and comprehensible format. Fourth, constituents of the DSLs (including the functors) are semantically grounded into a common ontology supporting syntactic and semantic correctness checks as well as code-generation capabilities. These techniques provide scalable, customizable, and reusable technical solutions for reoccurring multimodal processing tasks.