Thanks for the correction, but the key signature in a piece may be changing. Also, those with perfect pitch can tell the key signature from a short passage if there are few accidentals.
A common idea is doing modulation. It's usually an easy way to increase the excitement of the song. Most people can naturally hear and feel it, but maybe not have the vocabulary to describe what's happening.
Well, that's the thing -- when listening to any song, even if you know the pitches, it is impossible to know if there are accidentals or not depending on the phrase.
If a 10-bar phrase modulates in bars 2-5, and you listen to bars 2-5, you may conclude the piece is in key X when it's just a modulation from the overall key of Y.
This also implies the piece follows western common practice, or is even tonal in the first place. The difficulty of categorizing music is its ambiguity :)
The blog post criticizes existing software for their inaccuracy, but that is not their problem. The key signature of a particular passage may deviate from the overall key. I also wonder how they evaluate accuracy shown in [1] and what is the human accuracy in comparison.