Reading ASR and ACD: what call quality metrics actually tell you

Two figures dominate wholesale voice conversations: answer seizure ratio and average call duration. They are useful, they are easy to compute, and they are routinely used to argue that a route is good or bad. Read carefully, they say something real. Read on their own, they can point in exactly the wrong direction — and because they are the figures everyone has to hand, wrong conclusions drawn from them tend to travel quickly.

What each figure measures, and what it does not

Answer seizure ratio is the share of call attempts that are answered. Average call duration is how long answered calls last, on average. Both describe outcomes, not causes. A route does not have a quality figure the way a cable has a length: the number you see is produced jointly by the route, the traffic sent over it, and the people at the far end. That last point is what makes comparison difficult. The same route, carrying two different traffic profiles, will produce two different answer ratios, and neither of them is wrong. Several ordinary factors move the figure without saying anything at all about the route.

• →Traffic type — calls people expect are answered far more often than calls they do not, so appointment reminders and cold outreach will never show the same ratio.
• →Time of day — answer behaviour follows daily rhythms, so a comparison across different hours compares habits as much as routes.
• →Destination mix — aggregating several countries into one figure hides the single destination that is actually failing.
• →Number quality — calls to numbers no longer in service depress the ratio without reflecting on the route carrying them.

Duration cuts both ways

Average call duration is often read as a proxy for audio quality: if calls are short, something must sound wrong. Sometimes that is exactly right. Calls that end within the first few seconds, repeatedly, usually indicate that the person answering heard nothing — a one-way audio problem, or a long delay before the audio path opened. But short calls are also perfectly normal for some traffic, since a delivery notification does not need a long conversation. And a rising average is not automatically good news: it can mean calls are not being released cleanly when they should be, which tends to show up on the invoice before it shows up anywhere else.

Read them together, per destination, with release causes

One more distinction is worth keeping in mind: a figure computed over a handful of attempts is not a measurement, it is an anecdote. Small destinations produce dramatic-looking ratios that swing wildly from one day to the next, and acting on them wastes effort that a larger destination genuinely needs. Look at how many attempts sit behind a number before letting it change a decision. The practical approach is to look at both figures split by destination, alongside the release causes that accompany them. Release causes are the underused part of this picture: they tell you whether a call was rejected by the far network, abandoned by the caller, or dropped mid-conversation. Those are three different problems with three different owners, and no average will separate them for you. A route that looks poor in aggregate often turns out to be one destination behaving badly while the rest is healthy.

It also helps to agree what a comparison is for. Benchmarking two routes over the same period, with the same traffic, tells you something. Comparing this month against last month, when the campaign mix has changed, mostly tells you that the campaign mix has changed. Deciding which of the two you are doing before opening the figures prevents a good deal of fruitless argument. When we discuss quality with a partner, we look at the shape of the traffic before the averages, and per destination rather than in aggregate. It takes slightly longer than quoting a single number, and it produces conclusions that survive contact with the network.