PhoneticOperator, a voice recognition directory assistance product, handles a virtually
unlimited vocabulary through its speech recognition and directory search capabilities. The
directory, which can consist of up to a million names, is never subdivided or partitioned.
To provide error recovery capabilities to the caller and to find the correct record as
quickly as possible, Phonetic Systems patented software continuously searches the
entire directory throughout each stage of a callers dialog with the system.
Installation
PhoneticOperators installation is handled by Phonetic Systems themselves.
It is customized based on the needs of each individual client. However, we got a feel for
the installation process by setting up the hardware requirements and watching as a
representative from Phonetic Systems installed the software into NT Workstation 4.0. Since
we needed to install a Dialogic board, sound card and network card onto our system, we
used a 333MHZ Mitac industrial computer to assure that we would have enough slots to set
up the system adequately. (For more information about the Mitac industrial computer we
used, check the TMC
Labs Mitac review found in the September 1999 issue of Internet Telephony�
800-243-6002 or www.itmag.com.)
The biggest problem we had was obtaining enough RAM so we could load into the system
close to 400,000 names (four directories of varying size: 250,000 names, 100,000 names and
two smaller directories). We needed at least 256MB
of RAM to handle this. Once we were able to acquire this large amount of RAM, the software
installed quickly. It took less than 15 minutes to download all of the name directories
into the system, but every time we rebooted or updated the server, we had to wait those 12
to 15 minutes before we could continue testing PhoneticOperator. Fortunately, this does
not need to be done often, and a shadow database can be employed to help ensure that the
system is always running. If the server does go down, a program called Watchdog backs up
the server so that PhoneticOperator will still run.
Documentation
While browsing through the short users guide, we found the ingredients of a
good manual: a table of contents, constructive use of screenshots and tables, brevity and
a logical sequence of chapters. The users guide is satisfactory upon closer
inspection, as well. It is easy to follow, well thought out and most important, it is
helpful.
The telephone directories used for demonstrating speech-enabled searches are fine, but
the larger printed directories, such as the 250,000-name directory we looked at, are not
binded well because of their sheer bulk. Our suggestion here is to separate the large
directories into a group of smaller books and label each book by the first letter of the
last name (A-H, I-P and Q-Z, for example).
The help files do not disappoint either. They are concise and understandable as well as
context sensitive, displaying a short description about the topic that is currently
selected on the screen. If we have any complaint here, it is that these descriptions are
too short and may not completely help with a users problem.
Features
The following is a list of the major features of PhoneticOperator:
- Speech recognition technology handles up to one million directory records.
- Can use text-to-speech or can record a field using a human voice (but only 8 to 12
percent of names actually need to be prerecorded).
- Integrates with other telephony applications, existing PBXs and ACDs and other call
management systems.
- Searches entire directory continuously throughout a call.
- Makes directory updates immediately available for voice access.
- Error recovery capabilities used for more successful searches.
- Support for barge-in if using a Natural MicroSystems board.
Operational Testing
After importing four flat directories into our system, we tested the speech
recognition process by connecting a telephone to a simulator and asking PhoneticOperator
for specific names in the directory. PhoneticOperator searched through the 250,000-name
directory nearly as fast as the 15,000-name directory. That, in itself, was quite
impressive.
While PhoneticOperator immediately recognized some of the names we asked for, most of
the time it attempted to recognize the name by either asking for the city where the person
was located or for the spelling of the persons last or first name. If
PhoneticOperator was reasonably sure of the name we were requesting, it asked if that name
was correct. If it was, the dialog was finished. Normally, under real conditions instead
of lab conditions, it would route the call to that person once the name was confirmed. If
the name was not correct, it either asked us another question or said that it could not
find that person. Normally, this second option would result in routing us over to a live
operator. While the system recognized the spoken name much of the time, it is not 100
percent perfect, so therefore, an operator would still be needed to take those calls that
PhoneticOperator could not handle.
We then opened the Phonetic-Operator Administrator. The scripts and prompts form an
effective system that requires little work from the administrator, although you can still
add, delete or edit scripts as necessary. To edit, you must right-click on the name of the
script and change the conversational flow or settings of the script. The settings include
enabling barge-in, transferring calls to the operator, setting the maximum number of
questions and repetitions the system can ask and allowing for DTMF selections.
We examined other features as well. We added odd nicknames into a system that already
incorporated most of the common nicknames, such as Rich and Dick for Richard. We also
looked at the calendar, which acts as a basic scheduling tool that lists working hours and
employee holidays. One of our favorite features was the phonetic editor, which allowed us
to spell out difficult names by the way they sound or by their phonetic representation and
then incorporate the correct sound into the PhoneticOperator system.
Room For Improvement
When we first received Phonetic-Operator, we were immediately concerned about how
much memory it needed. No matter what the product is, 256 Megs of RAM is on the high side.
It would be nice to be able to add a large amount of names on a PC without needing quite
that much RAM.
Barge-in is a very important feature for speech recognition products because callers
can bypass unnecessary speech from the system. Being unable to use that feature with a
Dialogic board impedes the process. We realize that the hardware itself cannot handle
barge-in, but there can still be software made specifically for this purpose. Of course, a
Natural MicroSystems board could be used instead of a new software solution.
Since the directories for Phonetic-Operator are extremely large, it does not give you
the option to edit your name listings within the system without changing the parameters.
This option is easily accessible only for PhoneticAttendant, which is a product used for
systems of 2,000 directory names or less. Even though editing listings may often be
unnecessary for a system of PhoneticOperators magnitude, we think the option should
be easily accessible for this product as well as for the smaller one.
One last suggestion involves the spelling functionality of the system, which requires
that a user spell the entire first or last name. It would be easier and more expedient if
a caller was required to spell only part of the name to enable PhoneticOperator to search.
PhoneticOperator could identify the name based on the beginning of the spelling as well as
the other information the caller had given it before it had asked for the spelling.
Conclusion
Voice recognition directory assistance still has a ways to go before the
technology is perfected, but Phonetic Systems has a relatively reliable product in this
niche. For its sheer speed and ease of use, PhoneticOperator is definitely a system worth
considering for specific directory assistance purposes.
|