Latency (End-to-End Voice Latency): definition and how to read it
End-to-end voice latency is the delay between a caller finishing a sentence and hearing the agent begin to reply. It combines endpointing, recognition, the language model, synthesis and the telephony network.
Also known as: response latency, voice-to-voice latency, time to first audio
What adds to it
The pause is the sum of several parts: waiting to confirm the caller has stopped, transcribing the last words, language-model processing, any tool or knowledge lookup, speech synthesis and the audio's trip across the telephone network in both directions. Slow steps hide inside each other, so measuring only the model's time shows a flattering number the caller never experiences.
Why it matters
People expect conversational turns to start within a fraction of a second. Past roughly a second, callers start saying 'hello?' and talking over the agent, which causes recognition mistakes and abandoned calls. Lookups that call external systems are the usual cause of spikes and often deserve filler speech or caching.
Common mistake
Accepting a latency figure with no definition. Ask where it is measured (platform output or caller's ear), under what load, on which route, and whether it is an average or a high percentile. The slowest 5% of turns are the ones callers remember.
Where this shows up at Vaami
Vaami claims sub-300ms end-to-end response latency in production.
See it on a real call.
Hear how a Vaami voice agent handles this on a live conversation in your language and your use case.