Hi Andrewrona,
As you can see, this discussion can become very complex. As Gs5556 has pointed out, the most accurate transducer would reproduce the input energy without deviation (+-0db).
No speaker performs that well. Standard tests graph the performance and deviation with a recorded test at a standard input value, sweeping through the frequencys. This resulting ouput plot is the source information for the manufacturers statement that the speaker performs within +-3db over a specified frequency range.
Here is a site where you can listen to +-6db & +-3db changes in a note (half way down page - "Sound Example"):
http://www2.sfu.ca/sonic-studio/handbook/Decibel.html
Keep in mind that the 3db is the amount of sound level change the manufacturer has measured in your speaker's performance over the frequency range (from one frequency to another).
Here is a site that shows the relationship to db changes and the perception of loudness (Sound Pressure Level graph):
http://www.jimprice.com/prosound/db.htm
Keep in mind that the frequency where your manufacturer has measured the 3db variation may or may not be easily precieved.
The published specifications are from "on axis" measurments. They don't tell you anything about how the speaker sounds if you are moving in the room, or setting off axis. There will be significant changes off axis. Usually this off axis performance is one of the most variable qualities when comparing different speakers/manufacturers. Proper high frequency dispersion is one of the most crtical elements for good reproduction within the listening area. The specs can't help you to know how the speaker will sound in your room, because the room augments or subtracts sound pressure form various frequencys.
In my opinion if you have a speaker perfoming +-3db from 50-20Khz you have a very flat performance measurement. I would doubt that the db chages would be easily noticed by hearing. I have never seen a measurement with a completely flat line for the entire range.
Whether or not anyone likes the sound of the speaker can not be predicted by this published performance. Strong peaks in the response curve are more noticable than drop outs. And the high freqency dispersion is not indicated within this type of measurement.
Good luck,
Mark