International Links and VoIP: Bandwidth Planning for Global Offices

Running a global office means your team in London needs to talk to the team in Tokyo without sounding like they are underwater. Voice over Internet Protocol (VoIP) is a technology that transmits voice calls over IP networks instead of traditional phone lines. It cuts costs dramatically compared to legacy PSTN circuits. But if you get the math wrong on your international links, those savings vanish into dropped calls and robotic audio. The core problem isn't just buying more bandwidth; it's understanding how latency, jitter, and packet loss interact with your specific codec choices across oceans.

Key Takeaways

• Codecs dictate capacity: G.711 requires ~80 kbps per call, while compressed codecs like Opus or G.729 drop this to 30-40 kbps, but may struggle with high latency.
• Latency is physical: You cannot compress distance. A call between New York and London has a minimum round-trip time of 60-90 ms due to undersea fiber physics.
• Concurrency matters: Plan for peak usage. Use Erlang models to estimate that 5-20% of users will be on calls simultaneously during busy hours.
• QoS is non-negotiable: Without DiffServ markings (DSCP EF), VoIP traffic loses to file transfers and video streams, causing jitter and packet loss.
• Architecture affects load: Centralized architectures double traverse international links for local calls, wasting bandwidth and increasing latency.

The Physics of International Voice

When you move from local area networks (LAN) to wide area networks (WAN) spanning continents, bandwidth is no longer the only bottleneck. Speed of light is a hard limit. Data traveling through undersea fiber optic cables experiences propagation delay. For example, a signal between New York and London takes roughly 30-45 milliseconds one-way. Between London and Singapore, it jumps to 65-90 milliseconds. Add processing delays at routers, firewalls, and session border controllers (SBCs), and you quickly hit the thresholds defined by ITU-T G.114 is an international standard recommending one-way latency below 150 ms for acceptable voice quality.

If one-way latency exceeds 150 ms, conversations become difficult. People start talking over each other because the feedback loop is too slow. At 400 ms or more, normal conversation becomes nearly impossible. This is why satellite links, which introduce 500-700 ms round-trip times, are terrible for real-time voice despite having massive bandwidth. Your bandwidth planning must account for these physical realities. You can't buy your way out of latency, so you must optimize everything else-jitter and packet loss-to compensate.

Codec Selection: The Bandwidth vs. Quality Trade-off

Your choice of audio codec determines how much bandwidth each call consumes. A codec compresses audio data into packets. The three most common standards in enterprise environments are G.711, G.729, and Opus.

G.711 is the gold standard for clarity. It uses Pulse Code Modulation (PCM) and matches the quality of traditional landlines. However, it is heavy. With IP, UDP, and RTP headers added, each call consumes about 80-87 kbps. If you have 50 concurrent calls, that’s 4 Mbps dedicated solely to voice. G.729 compresses audio significantly, dropping the payload to 8 kbps. Total bandwidth per call falls to around 30-32 kbps. This was popular in the early 2000s for MPLS backbones where bandwidth was expensive. Today, Opus is a versatile audio codec standardized in 2012 that supports dynamic bitrates and handles packet loss better than older codecs. Many modern Unified Communications as a Service (UCaaS) providers default to Opus or similar adaptive codecs because they maintain quality even when network conditions fluctuate, typically settling around 40 kbps per call.

Calculating Concurrent Call Capacity

You don't need enough bandwidth for every employee to talk at once. That would be wildly inefficient. Instead, you use traffic engineering principles, specifically Erlang modeling, to estimate concurrency. In a typical office environment, 5% to 20% of users are on the phone during the busiest hour. Sales teams might hit 20%, while administrative staff might stay closer to 5%.

Let’s look at a concrete example. Imagine a sales office with 300 employees. Using a conservative 20% concurrency rate, you expect 60 simultaneous calls. If you use G.711 at 80 kbps per call, the calculation is simple: 60 calls × 80 kbps = 4,800 kbps, or 4.8 Mbps. If you switch to G.729 at 32 kbps, that drops to 1.92 Mbps.

However, never plan for exact numbers. Networks burst. Signaling traffic (SIP requests), encryption overhead (SRTP/TLS), and retransmissions add load. Industry best practice suggests adding a 20-30% headroom factor. So, for our 300-person office using G.711, you’d reserve roughly 6.2 Mbps (4.8 Mbps + 30%) for voice traffic alone. This ensures that when the sales team closes big deals during quarter-end rushes, the audio doesn’t break up.

Centralized vs. Distributed Architecture

How you route calls globally changes your bandwidth requirements drastically. There are two main approaches: centralized and distributed.

In a centralized model, all phones register to a single call control cluster or UCaaS region, often located in one data center. When an employee in Paris calls a local number in Paris, the call travels from Paris to the central hub (e.g., in New York), then back to Paris to connect to the PSTN. This "hairpinning" doubles the bandwidth consumption and adds significant latency. It simplifies management but punishes performance on international links.

A distributed model deploys regional Session Border Controllers (SBCs) or media gateways in key regions like Americas, EMEA, and APAC. Calls to local numbers break out locally. Only inter-office calls cross intercontinental links. This approach preserves bandwidth and reduces latency. Providers like Bandwidth.com and VoIP International emphasize the importance of local presence and emergency services compliance (like E911 in the US or 112 in Europe). A distributed architecture makes meeting these regulatory requirements easier because the call originates closer to the end user.

Quality of Service (QoS) Configuration

Having enough bandwidth is useless if your router treats a voice packet the same as a large file download. You must implement Quality of Service (QoS). QoS prioritizes traffic based on its type. For VoIP, you configure your edge routers and switches to mark voice packets with specific Differentiated Services Code Point (DSCP) values.

The standard marking for real-time media (RTP) is DSCP EF (Expedited Forwarding). Signaling traffic (SIP) is often marked as AF31 or CS3. When congestion occurs on your international link, the router processes EF-marked packets first. This prevents jitter-the variation in packet arrival time-which is the enemy of clear voice. ITU-T and Cisco guidelines recommend keeping jitter below 30 ms and packet loss below 1%. Ideally, aim for less than 0.1% loss for HD voice. Without QoS, a colleague uploading a design file could easily choke your VoIP traffic, leading to choppy audio and dropped calls.

Security Overhead and Encryption

Modern VoIP deployments encrypt both signaling and media. TLS secures SIP signaling, while Secure Real-Time Transport Protocol (SRTP) encrypts the audio stream. While essential for privacy and compliance, encryption adds computational load and header overhead. Each encrypted packet carries extra bytes. On top of that, if you tunnel VoIP traffic through a VPN (IPsec or SSL) to reach your corporate network, you add another layer of headers.

This overhead is usually small, around 5-15% additional bandwidth. However, in tight budget scenarios, it matters. More importantly, encryption can impact latency if your hardware lacks cryptographic acceleration engines. Ensure your SBCs and edge routers have sufficient CPU power to handle SRTP encryption without introducing processing delays that push you over the 150 ms latency threshold.

Monitoring and Analytics

Bandwidth planning isn't a one-time task. Network conditions change. ISPs upgrade routes; new applications compete for bandwidth. You need active monitoring. Use analytics dashboards provided by your UCaaS vendor (like Intermedia, RingCentral, or Zoom Phone) to track Mean Opinion Score (MOS), jitter, and latency per region. MOS scores range from 1 to 5, with 4.0+ considered excellent. If you see MOS dropping in specific regions, investigate the underlying internet connection. Often, the issue isn't the VoIP provider but the last-mile internet service at the branch office. Regularly review your QoS configurations and ensure that your SD-WAN policies are steering voice traffic over the lowest-latency paths available.