First I think that all of the answers above hold validity and should already give you some ideas about audio capability/ levels, etc.
To give you another simple implementation, I've linked to a post I wrote (including the electronics and code) on my blog here.. http://majordecibel.blogspot.com/2011/02/back-to-basics-reading-in-line-level.html
Basically, as mentioned, the arduino ADC reads 0-5V signals. The signal you are passing it would swing +/- about 0V. This means feeding it directly will clip the negative values; you'll just get the positive ones (not recommended, but you could go with it depending on your application).. An op-amp (also mentioned above will help - I think that one may also have a 2.5V bias on it, as I will describe..) if you tune it properly to scale the analog signal can also help to get maximum resolution (feeding in a larger analog signal that does not clip lets you take advantage of your full 10 bit sampling precision). Alternatively (simpler, but not quite as good) you can just turn the volume all the way up. You won't be using all 10 bits since it won't swing all the way up to 5V, but you will get enough to do something.
Anyways, without DSP related considerations to sampling/ anti-aliasing, etc, the key is that you want to bias the ADC pin to 2.5V using a voltage divider, and isolate that point from your audio source using a capacitor that will pass audio frequencies (4.7-10uF should be sufficient). That's three components; two resistors and a coupling capacitor. This lets you feed in both the positive and negative parts of the waveform. Input signals that are negative will swing the bias between 0 and 2.5V; Input signals that are positive will swing the bias between 2.5 and 5V.
In software, you can then subtract 512 to get your "negative" values. You can print this to the serial monitor to verify what's going on.
*Note: The code I included also has an "offset" value. You can calibrate this manually. You want to choose it such that after subtracting 512 you get 0 (the case when input signal is 0, so you just get the bias voltage). It may deviate from this if the voltage divider between the resistors is not exactly 2.5V (because there are inherence tolerances in the resistor values). Using larger resistors is fine - you are trying to establish a voltage bias; this helps to limit current and therefore power drawn from the supply... and given a fixed tolerance will result in less deviation of the bias point from Vcc/2 or 2.5V..
I could go a bit further, but I hope this will help, along with the posts above..