I really don't see the difference in two.
The difference is that for a Decidable language you can build a compiler that can always distinguish between valid and invalid "utterances". By contrast, if the language is only Recognizable, then there are invalid "utterances" that will cause a compiler to go into an infinite loop.
Note that by expressing this in terms of Turing machines, they are talking about what is theoretically possible; i.e. any theoretically possible compiler for the language, not just a specific one.
And the language "iff there is a Turing Machine ..." means that it is possible to formulate such a Turning Machine. It is not talking about all Turing Machines. A Turing Machine designed to (say) tell you if a number is odd obviously won't act as a language recognizer.
(If you don't understand why that is obviously so, you need to do some more background reading on the subject. I suggest you start with Wikipedia.)