Voice Mail Preview FAQ (Part 3)
This is the third part of the frequently-asked questions list that I’ve put together for Exchange 2010’s Voice Mail Preview feature. You may want to read Part 1 and Part 2 before reading this.
What happens if I get Voice Mail in more than one language?
In a UM Dial Plan, at most one language is used for Voice Mail Preview. Roughly speaking, a UM Dial Plan defines a collection of interconnected phones, in a customer site, that share the same numbering plan and for which UM is the voice mail system. Many companies have only one UM Dial Plan. UM will always attempt Voice Mail Preview, where enabled, in the default Dial Plan language (if this language is supported for Voice Mail Preview: see Part 2).
If a user has a mailbox that is UM-enabled in a US English Dial Plan, all their Voice Mail Previews will be created in US English. If, for example, they were to receive a Voice Mail in which the caller spoke in French, or German, or Turkish, the Preview would still be constructed on the assumption that the caller was speaking English. Almost certainly, every word of the Preview would be wrong.
Does the system “know” when a Voice Mail Preview is very inaccurate?
The system computes a confidence for every Preview that it creates. Loosely speaking, this measures how well the system thinks it did in matching the sounds in the recording with words, numbers and phrases in its repertoire. If the system found it easy to find matches, the confidence will be high. A higher confidence tends to be associated with a higher accuracy, but there is not an exact relationship.
The automatic speech recognition system cannot itself determine accuracy. If it could, it could “tune itself” and eventually become perfect. For the moment, accuracy can only be measured (in tests) with human intervention.
If the calculated confidence falls below a certain value, UM will indicate this by showing the words “Confidence is low” above the Preview. If these words are displayed, it is very unlikely that the Preview text will be accurate. Messages in which the Preview was constructed in the wrong language (see previous answer) will almost always be assigned a low confidence.
Does the system learn from its mistakes?
Automatic speech recognition software calculates a confidence in the preview, but it has no way to decide which words are wrong, and which are correct.
However, the system does try to learn from the environment to improve accuracy. For example, Exchange UM tries to match the caller’s telephone number (if provided) amongst the user’s Personal Contacts and in Active Directory. If the UM finds a match, it will include the name of the caller, along with its standard lists of names and words, when running automatic speech recognition on the voice recording.
Michael is a Lead Program Manager in the Exchange team, working on the Unified Messaging role of the product. He joined Microsoft in 2005, after 17 years in the software industry working on videotex, voice messaging, speech recognition services and miscellaneous other areas. His first contact with Exchange (as an ISV) was in 1993, at a time when it was still known as EMS, "Touchdown" and other code names.
In a previous life, Michael got a BA and D.Phil from the Univerity of York and spent several very happy years working as a researcher and lecturer at the University of California (San Francisco) and University College London. Among other things involved with the impecunious life of scientific research, he got involved with microcomputers. He taught himself Z80 assembler and C (on CP/M 1.4, if anyone out there remembers that). These days, he prefers to use C#, when the developers aren't looking.
