EMERTEC - FastAMD Documentation

Fast AMD is a new kind of digital AMD for use in data centres that put quality before quantity. Conventional AMD has been with us for a long time and its benefits are well understood…as to are its drawbacks. Fast-AMD from Magus approaches answer machine detection with a completely different way approach and overcomes these longstanding problems with old AMD.
Conventional AMD (sometimes called Cadence AMD) works making a good guess at what is making the noise that is coming down the line. The issues that discount it from being deployed in quality centric environments are principally two-fold. The first issue is that it can mistake live call recipients for answer machines and hang up on them; this is a false positive. Conventional AMD vendors will quote their false positive values at around 3%. Hmmm - maybe 20 years ago but with the subsequent domination of mobile/cell people just don’t answer calls in the same way they used to. Answer greetings are now on average longer and noisier, and cadence detection just cannot cope. 15% false positives are now the norm, and this is unacceptable in any environment
Secondly Conventional AMD takes too long. It takes at least 2 seconds for cadence AMD to begin to make a guess at what its connected too. This may not seem long, but that pause after an “Hello” is enough to alert the recipient that this is likely to be a call from a data centre. Forewarned, the call already starts on the wrong foot, even before your interviewer or agent has said a word.
Fast-AMD’s approach is to eradicate these two issues. Firstly, the Fast-AMD relies on finding a digital match for the incoming voice. As a result, there is a 0.0% rate of false positives. Its does not hang up on a connected live recipient. Secondly, if a digital match is possible, it can be made in around 0.5 seconds. That means Fast-AMD need to wait any longer than that before its transfers the call to your staff. Live call recipients are left unaware that there has been any delay.
However, you don’t get these benefits for free, there is a trade-off. Fast-AMD is implicitly more cautious that cadence AMD so it does not detect as many ‘real’ answer machines, but it still detects enough to provide significant performance gains. After you factor out the false positives, conventional AMD typically detects about 75% of all the answer machines and voice mails called. With Fast-AMD this number is down to about 60%, not as great but more than enough to improve the productivity of your staff, without annoying your call list.

magusAMD

The magusAMD (magus Answer Machine Detection) is an Asterisk Dialplan application designed to perform analysis of the in-band content of a call immediately after connection. It processes the media of the call and returns a best-guess as to the origin of the content.

It incorporates Asterisks own native AMD application features. It provides two more features over and above that provided by Asterisk’s native application.
1) Fax/modem tone detection
2) Fast Voice Mail message detection (fast-AMD).
The component functions, basic AMD, Fax detect and fast-AMD detect, can be run together or turned on/off independently.

The Fast-AMD feature

To perform its detection, fast-AMD uses a library of templates derived from the standard voice mail (VM) introductions. These are collected from the voice mail (VM) services of major national PSTN service providers and common PBXs. The content of a call is matched with these templates in real time. This matching may be done several times during the first fractions of a second. The matching is extremely fast (micro-second) so there is no lag between media collection and detection. This allows fast-AMD to determine an outcome on average within 0.4 seconds of the start of the call.

This extremely short detection time means that the “delay” experienced by live call parties is imperceptible. This is a huge benefit in countries where slow, traditional AMD is overused and the public are savvy to such practices. Many contact centres avoid using AMD simply because of the delay it places at the start of all connections. Fast-AMD removes this problem.

The make-up of digital templates and the matching algorithm mean that the probability of a false positive are around 2-30 (0.000000001); effectively there are no false positives. While basic AMD is “best guess”, fast-AMD is not. Templates are so unique that even the chances of the same individual re-producing a live voice segment to match a previous one are extremely remote (infinitesimal)

The avoidance of false positives is a major benefit in many countries where silent calls are limited by regulations (false positives come from both AMD and abandoned calls from predictive dialling). Typical false positive rates from traditional cadence based AMD are 5% - that means 5% of calls detected as AMs were in fact live call parties. Such a rate may disqualify the complementary use of predictive dialling. (It should be noted that we believe current false positive rates for traditional AMD are greater than 5%. With the increased use of mobile/cell over domestic lines, more individuals are combining personal and business usage. Live called parties are more often announcing with “Hello, Acme plumbing and heating, Kevin speaking”. Longer announcements mean more false positives)

Fast-AMD matches calls with a “known” message set. This means that it ignores domestic answer machines terminating on analogue lines. However in some countries (like the UK) “matchable” AM calls represent around 70% of all AM calls. Such limit are however outweighed by the effective removal of false positives and the speed with which detection is performed.

Library Management and Learning

The fast-AMD also has a dynamic and parallel learning feature. This layer of management allows templates from new, unmatched calls to be added to the library, however there is also “inactivity” algorithm that discards unused templates, and so prevents the build-up of unmatched templates. This is achieved by having 3 layers of library: undergraduate, graduate and postgraduate.

Postgraduates contains templates with a high frequency of matches, Graduates with a lower frequency, and Undergraduates are ‘new’ and have no matches yet. When a new call enters the system its generated templates are passed for comparison, first against the postgraduate set, then the graduate set and finally the undergraduate set.

When a call does not produce a match in the library, the templates from that call are added to the undergraduate layer. As it has already be passed against the undergraduate set, it must be unique. If, later, one of the undergraduate templates is matched with a new call, that template will get promoted to the graduate layer. If, after a configurable inactivity time (e.g. 3 days) an undergraduate is not matched, it will be removed.

Templates in the graduate layer are also subject to an inactivity time limit, but this is normally considerably higher (e.g. 21 days)

To support these learning options fast-AMD has 3 exclusive modes that can be run together or independently.
• Detect mode. This uses a set of pre-built templates for comparison with the current call media. It may also be set to use graduate templates.
• Template capture mode. This will capture a new set of templates based on the content of calls. These can be analysed in background to build or add to the library of templates used in detection mode.
• Media capture mode. This will save short media files (max 2 seconds) containing the initial part of call content. These may be used to qualify the captured templates as part of a manual graduation process

Templates and media files will be saved only when a fast-AMD detect match is not made. Saving of files may also be made conditional on a positive basic AMD detection (if that is running).

Installation

The installation is an rpm that places the binary and configuration files in various locations on the asterisk server. It also places a license control tool on the server.

Further information available from EMERTEC, call Linda Office: 01273 270270 Cell: 0776186180