Not sure what 'foobar' you're referring to there, but in general audioread does NOT read mp3 files correctly. As mentioned by MathWorks, this is due to limitations in the underlying operating system (Windows or Linux) that means audioread sometimes returns fewer samples than requested. Another limitation is that if you specify what sample indexes you want (via the 'samples' argument to audioread), and the sample range does not start at 1, you'll get different sample values at a given index than if the range starts at sample 1. Probably because of the way mp3 files are encoded, audioread doesn't (or can't) get the correct sample offset into the mp3 file.
I sometimes work with large (multi-gigabyte) sound files, so I have to process them in chunks. Being unable to read data starting at the correct sample offset into an mp3 file is a real headache. It turns out to be easiest just to convert mp3's to .wav files and work from there, even though the .wav files are much larger.
To answer your question, I'd use size(y) because that's what samples you can get.
