It seems that a standard such as CBOR (RFC 7049) would have been a better choice for several reasons:
- It's binary, so rather than having to base64-encode binary data for things like PCM samples to send to samplers, we could just send the data
- It's has very broad support, with implementations on most any language, just like JSON
- It's streamable
- It supports all the types JSON supports, and then some
- The type system is extensible, so types relevant to MIDI could be easily added
- It requires less CPU and memory to deal with, as parsing is simple.