Another approach for deciding if peak is significant might be to consider SNR in a box of given size centred around peak's frequency. I'm not sure if it is inherently better but might be worth testing and seems as pretty common approach.