Ok, then it's aparently the intermediate representation (what the parser feeds to the evaluator), which makes sense. The evaluator doesn't care, whether it comes from the parser or from constructs-to-c generated code.
Concerning the use case, an SBCL based solution would likely perform much better, though the Scheme VM of Common Music (https://ccrma.stanford.edu/software/snd/snd/s7.html) is also purely interpreted.