For any given network there will be some point at which the network capacity will become overloaded. This is often quoted as the maximum data throughput. For modern Ethernet this is often quoted as 1000Mbits/second.
However for networked voice communications network bandwidth will have an effect much below the saturation capacity of the network. The fundamental reason is that voice signals are realtime, continuous, and ordered. The significance of the realtime characteristic is that communications are nearly always two-way, requiring both parties to a particular communication to respond to each other. Any delay or other interference to the exchange of data will disrupt the communication experience. Equally speech tends to continuous, at least for significant periods of time, resulting in the continuous generation of voice data. Finally the order of speech data is fundamental in the ability to recognize the received voice data. It is not acceptable for the order of the speech data to become changed or broken-up.
ASTi would recommend from experience that a practical limit on network load is 30% for clean voice communications. This translates to a loading of 3000 active voice streams for a 100Mbs Ethernet which is quite a considerable number. Above the 30% load point collisions on the wire begin to occur more frequently causing re-transmission delays to occur, this leading to delayed and out-of-order packet reception, and ultimately broken up voice reception.
The above analysis is based on Half-Duplex Ethernet, however similar networking principals can be used to make conclusions about full duplex Ethernet, ATM or other underlying Data Link and Physical communications layers associated with a network segment.