We can't know if analog volume changes (external speakers, etc), but if the OS changes the volume we generally can find out and should reset the AEC in most cases (especially on volume-up). We may want to experiment and make it depend on the amount of volume change as well. After a volume change, the AEC training is at best compromised, and typically the AEC seems to need 5-10 seconds after a volume change to regain its previous quality. The belief is that there will be less artifacts (and faster training) if we force it to reset.
Actually, if it's just a slight volume change, in theory the best option is *not* reset the AEC, but somehow let it know it needs to do fast adaptation like when it's been reset. The only exception is when the volume gets reduced by more than about 6 dB (in which case the AEC would be adding echo). Of course, this is the theory, so it would need to be confirmed in practice. Just saying that a reset is not obviously better.