First of all I suggest you set up the system so that the voice traffic is first offered to the voice only group overflowing onto the common group. This will bemore efficient than offerring the voice to the common group first.
If the ammount of data traffic is say 50% or more of the voice traffic we could use ErlangB for sizing ignoring the peakedness of the overflow voice traffic. Otherwise we need amore sophisticated method to estimate blocking performance.
Therefore I suggest as a first shot to dimension voice only circuits to a blocking of 10% using Erlang B. Then add the volume of overflow voice to the volume of switched data traffic and again dimension to a GOS of 10%. This will get you in the correct ball park but will underestimate the number of circuits required if the volume of overflow voice is high relative to the data traffic. In this case you need to use more sophisticated methods. Wilkinsons equivalent random traffic method comes to mind.
There are two methods commonly used to analyse such systems (assuming random traffic and exponentially distributed holding times). The first is to use a 2-dimensional Markovian system analysis. The second is to use simulation.
Send me an Email on firstname.lastname@example.org if you want further information.