In the past, all voice networks were built using a centralized architecture in which dumb endpoints (telephones) were controlled by centralized switches. Although this model worked well for basic telephony services, it mandated a trade-off between simplified management and endpoint and service innovation.

One of the benefits of VoIP technology is that it allows networks to be built using either a centralized or a distributed architecture. This flexibility allows companies to build networks characterized by both simplified management and endpoint innovation, depending on the protocol used.

In general, centralized architectures are associated with MGCP and H.248/MEGACO protocols. These protocols were designed for a centralized device—called a media gateway controller or call agent—that handles switching logic and call control. The centralized device talks to media gateways, which route and transmit the audio/media portion of the calls (the actual voice information).

In centralized architectures, the network intelligence is centralized and endpoints are relatively dumb (with limited or no native features). Although most centralized VoIP architectures use MGCP or H.248/MEGACO protocols, it is also possible to build SIP or H.323 networks in a centralized fashion using back-to-back user agents (B2BUAs) or gatekeeper routed call signaling (GKRCS), respectively.

Advocates of centralized VoIP architectures favor this model because it centralizes management, provisioning, and call control. It simplifies call flows for replicating legacy voice features. And it is easy for legacy voice engineers to understand. Critics of centralized architectures claim that it stifles innovation of endpoint features and that it will become a hindrance when building VoIP services that move beyond legacy voice features.

Distributed architectures are associated with H.323 and SIP protocols. These protocols allow network intelligence to be distributed between endpoints and call-control devices. Intelligence in this instance refers to call state, calling features, call routing, provisioning, billing, or any other aspect of call handling. The endpoints can be VoIP gateways, IP phones, media servers, or any device that can initiate and terminate a VoIP call. The call-control devices are called gatekeepers in an H.323 network, and proxy or redirect servers in a SIP network.

Advocates of distributed architectures favor this model because of its flexibility. It allows VoIP applications to be treated like any other distributed IP application, and it allows the flexibility to add intelligence to either endpoints or call-control devices, depending on the business and technology requirements of the network. Distributed architectures are usually well understood by engineers who run IP data networks. Critics of distributed architectures commonly point to the existing PSTN infrastructure as the only reference model that should be used when trying to replicate legacy voice services. They also note that distributed networks tend to be more complex.