The problem is with option "Duck when outputing speech and sounds", which seems to be the default.
When playing music for example, the sound is correctly ducked when NVDA speaks, but forgets to get back to its original level when NVDA stops, wich probably is the purpose of this option.