For Australian businesses deploying call AI, local hosting is a fundamental requirement. The physics of international data transmission create unavoidable delays that make overseas-hosted voice AI feel laggy and frustrating. Under 600ms feels natural. Over 900ms? People hang up.
The Physics Problem You Can't Engineer Around
The laws of physics dictate that data can't travel faster than light. Light moves through fibre optic cables at roughly 200,000km per second. Sydney to San Francisco is approximately 12,000km.
So the theoretical minimum round-trip time (aka ping) to the US West Coast: 120ms
Typical round-trip time (ping) with network routing: 150-200ms
That's before anything else happens. Before phone network hops. Before speech recognition. Before the AI thinks. Before text-to-speech. Just the travel time of data in an optic fibre cable under the Pacific Ocean.
And it happens potentially multiple times per turn within a conversation.
What Latency Actually Does to Calls
We've tested this extensively with Australian users. Their feedback is consistent:
"There's this weird pause after I speak. Makes me think it didn't hear me."
"I kept talking because I thought the call dropped."
"It's like talking to someone on a bad international line."
"We kept interrupting each other."
"I gave up and called back to speak to a real person."
Remember, these are just ordinary people trying to get something done. When latency creates friction, they disengage.
Technology should serve our needs and make things easier for us, not get in the way. What Is Latency in AI Voice Calls?
The Breakdown by Response Time
Under 400ms: Natural conversation. Turn-taking feels right. Speakers don't notice delay. High completion rates.
400-600ms: Noticeable but tolerable. Slight hesitation. Speakers adapt by pausing longer. Some frustration accumulates.
600-800ms: Problematic. Clear lag between exchanges. People talk over each other. Repeated misunderstandings. Completion rates drop significantly.
Over 800ms: Broken. Conversation rhythm impossible. Frequent overlapping speech. High abandonment. Callers assume the system is broken.
The Real Numbers
Let me show you the maths:
Voxworks (Australian-hosted) typical latency:
- Network: 15ms
- Voice Activity Detection: 50ms
- Speech-to-Text: 150ms
- LLM processing: 300ms