You state "the Windows kernel can be used for audio recording, gaming or anything without a delay".
There is in fact a delay, so what you really mean is "without a delay that is perceptible or troublesome to humans". The delay is unavoidable, but the size of the delay is what determines a system's suitability for "interactive" use.
There has been a lot of work put into low-latency audio frameworks - for example Windows' ASIO, and Linux's JACK. These typically involve removing layers of abstraction, reducing buffer sizes, etc...
Media playback is just about synchronising the audio and video elements, the buffers can be huge so long as the playout is coordinated correctly.
Gaming and Pro Audio Recording require streams to be built on-the-fly depending on external factors like user input.
The topic of "Real-Time Computing" has no real place when discussing human interaction with systems - we are very lenient and our brain can handle relatively large offsets between our senses - for example audio and video.
Real-Time systems provide guarantees - for example if a trigger input occurs, computation will complete and output will occur within n microseconds. This is crucial for correct operation of many systems - often with life or death consequences.
The significant additional effort involved in providing such guarantees - in design, implementation and verification - is simply not justifiable to make sure that your game feels immersive - not to mention that such systems are often orders of magnitude simpler as a side effect of assuring the guarantees are met.