Answering Machine Detection (Amd) - AudioCodes Mediant 4000 SBC User Manual

User's Manual
Call-ID: 11923@10.33.2.53
CSeq: 1 INVITE
Contact: <sip:100@10.33.2.53>
X- Detect: Response=AMD,CPT
Content-Type: Application/X-Detect
Content-Length: xxx
Type = CPT
Subtype = Beep
14.5

Answering Machine Detection (AMD)

The device's Answering Machine Detection (AMD) feature can detect whether an outbound
call has been answered by a human (including fax) or an answering machine. The device
analyzes the sound (speech) patterns received in the first few seconds of the call to
determine whether a human (live person) or machine has answered the call. Typically,
when a human answers the call, there is a short "hello ..." followed by silence to wait for the
other party to respond. In contrast, when an answering machine answers the call, there is
constant speech (answering message) followed by a beep to leave a voice-mail message.
When the device detects what answered the call (human or machine), it can notify this
detection type to, for example, a third-party application server used for automatic dialing
applications. The X-Detect SIP header is used for requesting event detection and
notification. For more information, see ''Event Detection and Notification using X-Detect
Header'' on page 187. The device can also detect beeps played by an answering machine
at the end of its greeting message. For more information, see ''Detecting Answering
Machine Beeps'' on page 188.
The device's default AMD feature is based on voice detection for North American English
(see note below). It uses AudioCodes' sophisticated speech detection algorithms which are
based on hundreds of real-life recordings of answered calls by live voice and answering
machines in English. The algorithms are used to detect whether it's human or machine
based on voice and silence duration as well as speech patterns. The algorithms of the
language-based recordings are compiled into a file called AMD Sensitivity. This file is
provided by default, pre-installed on the device.
Note:
pattern and silence duration, the language on which the detection algorithm is based,
is in most cases not important as these factors are similar across most languages.
Therefore, the default, pre-installed AMD Sensitivity file, which is based on North
American English, may suffice your deployment even if the device is located in a
region where a language other than English is used.
However, if (despite the information stated in the note above) you wish to implement AMD
in a different language or region, or if you wish to fine-tune the default AMD algorithms to
suit your specific deployment, please contact your AudioCodes sales representative for
more information on this service. You will be typically required to provide AudioCodes with
a database of recorded voices (calls) in the language on which the device's AMD feature
can base its voice detector algorithms. The data needed for an accurate calibration should
be recorded under the following guidelines:

Statistical accuracy: The number of recorded calls should be as high as possible (at
least 100) and varied. The calls must be made to different people. The calls must be
made in the specific location in which the device's AMD feature is to operate.

Real-life recording: The recordings should simulate real-life answering of a called
person picking up the phone, and without the caller speaking.

Normal environment interferences: The environment in which the recordings are done
should simulate real-life scenarios, in other words, not sterile but not too noisy either.
Version 7.2
As the main factor (algorithm) for detecting human and machine is the voice
191
14. Media
Mediant 4000 SBC