Demo calls always happen in a quiet room, on a good line, with someone giving the call their full attention. Real calls are not like that: the customer is in a car, in a supermarket, a child is crying, the line is weak.
The quality of a voice agent is measured not by how fluently it speaks in good conditions but by what it does in bad ones.
Three separate problems, three separate fixes
These cannot be lumped under one "audio quality" heading; their causes and remedies differ.
Background noise. The customer is not speaking but the environment is producing sound. Risk: the agent mistakes it for speech and interrupts or tries to answer.
Silence. The customer is thinking, looking something up, or did not hear. Risk: the agent treats silence as "they finished" and moves on — or waits forever.
Drops. The line cuts for a second or fails entirely. Risk: the agent does not notice and keeps talking into nothing.
The right behaviour in noise
What stops an agent from mistaking noise for speech is an input gate: a threshold deciding whether incoming sound is really speech. Set it too low and the agent constantly interrupts; too high and the agent talks over the customer.
The correct value varies per deployment and is tuned by ear: listening to real recordings and noting where the agent came in early beats any theoretical number.
A special case: acknowledgements like "mm-hm", "yes", "okay". These are speech but not a request to take the floor. An agent that stops and waits for them stalls the conversation artificially.
The right behaviour in silence
Silence handling rests on two parameters: how long to wait and what to say when the wait ends.
- Wait time should depend on context. If the agent asked a question it should wait longer; the customer may be looking something up.
- The line spoken after the wait should repeat the question, not ask "are you there?". "Are you there" resets the call; repeating the question advances it.
- After a second silence the agent should move toward closing rather than waiting for a third.
The right behaviour on a drop
For a full drop the action is clear: mark the record "line dropped" and return it to the call queue. What matters here is not counting a dropped call as "connected".
Partial drops are sneakier: audio cuts for a few seconds then returns. The customer says "hello? hello?". If the agent carried on with its sentence meanwhile, neither side hears the other. The right behaviour is for the agent to restart its sentence after a prolonged loss of audio and ask for a short confirmation.
The silent failure: an agent that never speaks
The worst scenario is an open line with no sound from the agent at all. The customer waits on an empty line and hangs up. This is not a conversational error but a technical fault, and it shows in the records as: short call duration, no speech on the agent side.
A simple rule catches it: raise an alert if no audio is produced on the agent side within a few seconds of the line opening. The difference between noticing a silent fault in seconds versus hours is dozens of lost calls.
Measurement
Three numbers deserve regular tracking: average call duration, the share of calls ending under ten seconds, and the count of calls where the agent never spoke. Together they warn you about quality degradation before your team does.
Tomorrow we look at the list itself: weeding out dead numbers.