Call silence and interruptions: what they really show

Short answer

Silence, hold and interruptions are neither good nor bad on their own. They are signals that something is happening in the call: the agent is searching a system, the customer is looking for a document, two people started talking at once. What a signal means only becomes clear in context — the call type, what was said before the silence and what happened after it.

So the right use of these metrics is not a rule like "silence over 30 seconds costs points". The right use is to turn them into a way of selecting calls for review: find calls with unusually long silence, then look at their content.

Three signals, three different sources

1. 1HoldThe agent putting the customer on hold with the hold button. This is usually logged as a separate event by the telephony system and is the most reliably measured signal.
2. 2SilenceThe line is open, but nobody is speaking. Measuring it means separating speech from non-speech in the audio itself. A transcript's text does not show silence length reliably.
3. 3Interruptions and overlapBoth sides speaking at once. Seeing it requires each side's speech separated in time — ideally on separate audio channels.

AWS's Contact Lens documentation presents talk time and non-talk time as separate metrics for agent and customer. That is one platform's approach; other tools may not have these metrics at all, or may compute them differently.

When silence is not a problem

• The agent is looking up the customer's details and said so first: "One moment, I'm finding your order"
• The customer is looking for a card number, contract number or document
• The agent has started the after-call note during the conversation
• The customer is thinking about an offer — in a sales call a short silence is often the moment of decision
• A technical fault: audio dropped on one side while the line stayed open

The last case matters: an audio drop is not the agent's fault, but it can show up in a silence metric as the agent's silence.

When silence is a problem

• The agent goes quiet for a long time without a word — the customer thinks the line dropped
• One call has several long silences, all around the same question — the agent does not know the answer
• The hold was not explained to the customer, or lasted much longer than the time stated
• After the silence the customer's words change: "Hello? Is anyone there?"
• Every agent has long silences on the same type of question — the problem is in the knowledge base or the system's speed

How to read interruptions

Not every overlap is rude. In live speech, short acknowledgements like "yes" or "I see" are normal and tell the customer they are being heard. The problem lies elsewhere:

1. 1NormalShort acknowledgements; both sides starting at once and one of them stopping; the customer helping the agent finish a sentence.
2. 2Needs attentionThe agent often cuts across the customer's sentence to give their answer; the customer has to repeat the same thing several times.
3. 3SeriousWhile an upset customer is speaking, the agent talks over them and raises their voice; the customer says "you're not listening to me".

Counting interruptions has technical preconditions too: when two voices mix in a mono recording, the system can get the speaker wrong. In those cases an interruption count is not reliable.

Illustrative example: an interpretation matrix

The matrix below is an illustrative example. For each signal it answers "what should be checked" — a review step, not a score.

1. 140-second silence, agent said "let me check" firstCheck: which system was the search in? If it is the same for every agent, it is a process issue.
2. 240-second silence, nothing saidCheck: was there an audio drop? If not, remind the agent of the "say it before you go quiet" rule.
3. 33-minute hold, "one minute" was promisedCheck: did the agent explain the situation on returning? A gap between stated and real time is a behaviour topic.
4. 4Frequent overlap, calm customerCheck: are most of them short acknowledgements? If so, there is no problem.
5. 5Frequent overlap, upset customerCheck: is the agent answering without listening? That is a de-escalation coaching topic.

Do not turn the metric into a score

Setting a threshold for silence or interruptions and deducting points automatically is tempting, because the number looks objective. It is a mistake, for three reasons:

• The same number means different things on different call types — a long search is normal in technical support, not on a quick information call
• The signal can have a technical cause — an audio drop, a poor recording, speakers separated wrongly
• Agents learn the score: they fill silence with filler words and avoid the hold button

A better approach: use the signals as a selection criterion for review, and do the evaluation with behavioural criteria — for example, "the reason and expected time were stated before the hold".

Criteria you can write into a call standard

• The agent stated the reason and expected time before putting the customer on hold
• On returning from hold, the agent explained the result or the next step
• During a long search, the agent updated the customer periodically
• The agent did not cut across the customer and let them finish their sentence
• The customer did not have to repeat the same thing

These criteria are observable behaviour, not time measurements. They can be checked against the words in the transcript.

Look at the team, not only the agent

The most valuable use of the silence signal is often not about an individual agent. If the whole team has long silences and holds on the same contact type, the cause is not the agents but the environment they work in.

1. 1Group by contact typeSplit the calls with long holds by topic: tariff question, order status, technical fault, refund.
2. 2Find the topic that repeats mostIf one topic accounts for most of the long holds, that is where agents cannot find the answer quickly.
3. 3Separate the root causeIs the answer missing from the knowledge base? Is a system slow to open? Is an answer awaited from another team? Each goes to a different owner.
4. 4Measure again after the fixAfter the knowledge base is updated or the process changes, check with the same method whether holds on that topic got shorter.

This protects agents from being labelled "slow" and shows management the real problem: information that is hard to reach.

Limits

• A transcript's text does not show the exact length of a silence or hold; that needs audio timing analysis or telephony events
• In a mono recording, speakers separated wrongly can distort interruption counts
• A poor or broken recording can look like silence and must not count against the agent
• One platform's metric cannot be compared directly with another platform's

What Vexvon Audio Analyzer offers

Vexvon Audio Analyzer does not produce timing metrics for silence, hold or interruptions. Its strength is the content around the signal:

• Each transcript line is attributed to the agent, the customer or — when the system is not confident — "unknown"; in a stereo recording the sides are separated by channel
• You write the behavioural criteria above into your call standard, and the system judges them on every call as met, partial, missed or not applicable
• Each judgement comes with the transcript lines behind it — for example, the sentence the agent said before the hold

Pull the calls with long holds from your telephony, upload them, and you get an evidence-based answer to "was the hold explained?". More: Vexvon Audio Analyzer. How to read overall call length is covered in AHT vs call quality.

First step

Take the 10 calls with the longest holds from your telephony report and ask two questions of each: was the hold explained to the customer, and did the agent give the result on returning? The answers will show which criterion to add to your call standard. This article belongs to the support quality and coaching section. To try it on your own recordings, get in touch.

Further reading on this topic: audio analytics vs transcript analytics, speaker diarization call center.