Showing results for 
Search instead for 
Did you mean: 

UCCE CPA Answering Machine Detection - voice detected as voice machine

Slavik Bialik
Rising star
Rising star

Hi guys,

You'll probably think it is a joke, but it is real and I also tested it.

It looks like the CPA has an issue when a person, male, with Russian accent is answering the dialer call,

the CPA is detecting him as a voice mail and disconnects the call in their face.

I even tested it on my own when our client complained about it, I happen to have a Russian accent, and few times it detected me as a voice mail,

and the call was disconnected. Of course I checked in the UCCE DB, in Dialer Detail table, I've got CallResult = 12, meaning "Answering Machine Detected".

Funny thing is that when I distorted my voice a little bit, the call answered and referred to an agent.


We came to conclusion that it has nothing to do with accents, because the algorithm is working in other ways. But still the issue occurs when voice calls are detected as voice machines. 

I'm currently using UCCE 10.5 (MR2).

I also read this article:

Answering Machine Detection Algorithm

It's a very old one. I'm not even sure if it still the algorithm, or Cisco changed it. But if not, I have no idea how can I try and tweak the registry and values of the timers so it can sort the issue above, if possible.

Looks like Cisco doesn't like Russian males, heheh ;)

Anyway, does anyone have an idea how to sort it out, or tweak it in a way that will reduce those disconnections?



10 Replies 10

Edward Umansky

The algorithm relies 100% on voice volume and time, nothing specific to accents. It could be with the russian accent your volume is different or greeting is longer. You could tweak the thresholds so that live voice detect allows slightly longer greetings.

However, in my experience it is *very* easy to go astray trying to perfect things in testing. I strongly recommend that you do not under any circumstances tweak the algorithm based on your development testing. At best you'll improve things slightly for the few numbers your testing with, while making everything worse for actual customers. The only recommended way to tune the dialer is take it live with a pilot, enable CPA recording, then evaluate a large chunk of recordings to identify your accuracy % and note any consistent issues. From there you may have the info you need to determine whether tuning is needed and how you might tune it.

Hi Edward, thanks for the reply.

Yesterday, I enabled the CPA recording, and I took 50 recordings to analyze.

After analyzing 50 of them, I understood that the accent thing is nonsense. Actually, that's the input I got from our client (that's what he thought) and that's a feeling I got after testing it 3 times by myself (my calls got disconnected).

So, out of 50 recordings, 9 of them disconnected. Which is very bad statistics.

2 of them that I can remember, I analyzed and found out that they're starting with some kind of a beep, but after analyzing in GoldWave, I found out that this is a small chunk of ringback tone.

Probably the call was using 183 instead of 180, and therefore in some way the ringback tone was mixed into the call of the dialer. I don't have the SIP messages of those calls, so it's a fair guess.

Anyway, it made me think, that maybe the algorithm counted it as some kind of a beep and thought It's a voice machine. Is it possible?

The rest of the calls, I couldn't understand why and how the dialer thought those are voice machines, and it disconnected them also.

I'll post some of them on Sunday + the campaign configuration, I hope that someone maybe can take a look and understand how I can tweak it to be a little bit better. Because 9 out of 50 calls, it's bad.

Thanks again.


To be clear, you found 50 calls that you identified as voice. Out of those, the dialer identified 9 as answering machine? Did you confirm that indeed the 9 had result 12 and the remainder had result 10?

OK, I explained it wrong. I took a total of 50 files to analyze.

22 of them were real voice machines. The rest (28) I classified as voice calls.

It means that out of 28 voice calls, 9 of them disconnected. heh, now the statistics even worst.

And yeah, every call I tested, I used its call GUID to check in the Dialer_Detail table in the DB to see if the result is 10 or 12 for each call, or something else. In that, I'm 100% sure.

You can see in this picture my current CPA configuration for my campaigns:

Campaign Configurations


Ah ok, so out of 28 calls you identified as voice, 9 of them were marked as result 12? That does sound like something is wrong.


I can post on Sunday those calls as a .wav files, maybe it can help us understand why the algorithm treats them as a voice machines.

Unless you know some recommended configurations for the campaign CPA settings that aren't the default ones?

Ya I can took a look at the files. Default should not be causing issues to that degree. It may be a good idea to add another random 50 files to your analysis. I know how annoying going through them could be but may make it easier to detect a pattern.

Hi Edward, thanks for the help.

I'm attaching the .zip file that is containing:

1. Excel file with all of the results (pay attention to those that aren't in green color).

2. Folder with all of the recordings that had issues.

There are 2 more recordings in there, except the 9 I was talking about. There was one call that identified as fax, and of them was an actual voice machine but detected as voice.

You can see in the excel results file all my comments that I put for each call. Probably some lucky guesses, but can't be sure if it's correct or not. Some of them, I couldn't even guess what's wrong.

Thanks again! I really hope you'll see something in those.

So looking at these, it seems like most of them are due to a lot of background noise or bad connections. You can try adjusting the noise / volume thresholds to see if it helps. Hard to say what's going on otherwise, might be a good idea to get more samplings from various times of day to see if these noise issues are consistent.

Hi Edwards,

Thank you for reviewing my recordings!

Can I ask how do you notice that there are lot of background noises? So when I'll try to sample some more recordings I will be able to see that also.

Thanks a lot!

Getting Started

Find answers to your questions by entering keywords or phrases in the Search bar above. New here? Use these resources to familiarize yourself with the community:

Recognize Your Peers