Accuracy changes with audio length

Is it normal that model accuracy increases when the audio is split into short segments and why? how can I minimize this problem when working with whole and long audios?
I used aspire (en) and zamia (fr) and both have this problem