One Number, Follow Me
Not so long ago, the venerable business card typically listed just one telephone number. Increasingly, though, today's business card is cluttered with different telephone numbers for direct line, cellular phone, pager, and sometimes even home office. All those lines - while theoretically keeping you connected - are a nuisance for customers and business associates who have to figure out which number to phone at any given time.
Now, however, new telephone services are putting some simplicity back into callers' lives by giving them one number to phone and letting technology automatically switch them to the right number. One-number - follow-me services are particularly important in today's mobile business world, where being connected can spell the difference between success and mediocrity.
To use these one-number follow-me services, you typically specify all of your different phone numbers to which your service should automatically forward calls. The service lets you prioritize the numbers.
There's just one problem, however. Many of the services are programmed with a touch-tone keypad and a series of voice menus that prompt you through the set up, such as "Press 1 for personal options, Press 2 for forwarding options, etc." It can be very difficult for customers to remember which DTMF signal means which command, which can slice into the productivity gain originally antici-pated.
AccessLine Technologies, Inc. of Bellevue, Wash., however, has hit upon a solution which uses automatic speech recognition (ASR) technologies from Applied Language Technologies (ALTech) of Cambridge, MA and Dialogic Corp. of Parsippany, NJ that lets you control and program the service with spoken commands. Specifically, AccessLine's One Person One Number(r) service uses ALTech's SpeechWorks(tm) phonetic automatic speech recognition (ASR) software and Custom Context DialogModules(tm) running on Dialogic's Antares 2000/50 DSP speech platform.
"Our One Person One Number(r) service may eventually obsolete the touch-tone keypad in one-number services, much like DTMF has rendered obsolete the pulse code dialing of the rotary telephone," said Fred Epler, senior vice president of advanced technology at AccessLine. "The combination of ALTech's ASR software solution and Dialogic's DSP platform enabled us to build a natural, conversational spoken interface for programming and delivering our one-number service."
The ALTech software combines three important pieces in building a conversational command structure: a large vocabulary, speaker independence, and continuous automatic speech recognition.
• Speaker independence employs a special speech recognition algorithm that can recognize words spoken by a broad base of the population, eliminating the need to "train" the speech recognition system to one voice.
• Continuous recognition refers to the ability of the software to "listen" to continuous speech without requiring the separation of each word by a fixed amount of "quiet" time or forcing the user to wait for a beep before pronouncing each word in the sentence.
• A large phonetic vocabulary uses phonetic representation of the words in that vocabulary; the vocabulary is built by simply typing in the desired words.
This combination of technologies creates a natural, conversational interface for programming the system. For example, a subscriber can tell the system to "Forward all of my calls to my house for the next two hours," or "Forward all urgent calls to 555-1234 until 7 p.m." ALTech's innovative software gleans the necessary pieces of information from the user's conversation and commands the system to perform the desired action. Further, the real-time vocabulary generation and large vocabulary capabilities of SpeechWorks support the recognition of all of the intuitive command words, proper names, and directories needed to implement the service.
AccessLine's One Person One Number(r) service is built on their AIS (Advanced Intelligent Service) platform that resides in a central office or switch environment and is connected to the Public Switched Telephone Network (PSTN) via T1, E1, ISDN PRI, or SS7 lines. Thus, it's critical that the ASR used be robust enough to recognize speech from male and female subscribers over both cellular and landline networks. ALTech's recognition engine typically achieves raw accuracy scores of 95% and greater.
However, the real challenge in implementing a large vocabulary ASR application in the telephone network is not in improving the raw accuracy scores to 97% or even 98%, but more in the human factors engineering. Since 100% raw accuracy is not possible even for human listeners, the real issue is how to design an ASR system that recovers gracefully from a recognition error to ultimately achieve an overall task completion rate of 100 percent.
ALTech's strategy is to mimic the process that humans go through: Apologizing for the miss and either requesting that the subscriber repeat the command or prompting the subscriber for additional information, including a request for spelling a command or other word in question.
In developing large vocabulary systems, ALTech has found that even the way prompts are worded can affect overall ASR accuracy and subscriber satisfaction. By carefully wording the prompts, ALTech's solution can lead the user to respond to the prompt in a predictable fashion, thereby easing the burden on the recognition engine. For example, a prompt that says, "That line is busy. Do you want to try again in 90 seconds?" predictably generates a yes or no response; conversely, a prompt that says, "That line is busy. When do you want to try again?" can result in a collection of unpredictable responses, ranging from "Now," to "In 15 minutes, " to "Forget it ... I can't wait." This broad variation of answers challenges the accuracy capabilities of the recognition engine.
AccessLine's One Person One Number(r) service was built using an active vocabulary of over 1,000 words, including the capability of recognizing natural numbers. For example, it recognizes "one eight hundred, five, five, five, five thousand" for 1-800-555-5000, as well as "one, eight, zero, zero, five, five, five, five, zero, zero, zero." The same is true for dates and times.
For experienced users, AccessLine's One Person One Number(r) service deploys echo cancellation algorithms that allow the user to respond to the prompt at any time, which lets subscribers move onto the next action without waiting for the end of the prompt's message or a beep.
In selecting technologies for its speech-enabled control system, AccessLine specifically selected ASR solutions which were modular and scalable, allowing customers to rapidly expand capacity as the subscriber base increases. Both ALTech's ASR software and Dialogic's Antares 2000/50 DSP speech platform met these important specifications, enabling ALTech to offer three configurations - ranging from 1 to 500,000 subscribers - using the same technologies. Increasing the number of system ports requires the simple addition of more Antares boards.
Based on standard digital-signal processing chips, the Antares platform supports the SCBus protocol and multiple ASR algorithms used in offerings from Voice Control Systems, Lernout & Hauspie, PureSpeech, Nuance Communications, Dragon Systems, Syrinx Speech Systems, CNET, CSELT, and Telefonica. The Antares board also supports text-to-speech and speaker verification services, which AccessLine plans to use in its soon-to-be-released Internet subscriber interface unveiled at PCS '96 in September.
Another factor in AccessLine's decision to work with Dialogic and ALTech in creating its system was time to market. By running ALTech's high level DialogModule(tm) application building blocks on the Antares board, AccessLine was able to build its voice enabled interface in less than three months. The DialogModules(tm) were developed by ALTech to accomplish specific recognition tasks within an overall application, such as capturing a yes/no response, capturing a phone number, listening for voicemail command words, specifying a list of names and phone numbers, and capturing address information.
Thus, the DialogModules(tm) offer complete functionality out of the box, streamlining the overall integration of ASR in an application. The cost of deploying speech recognition in this application represents less than 10% of the total system cost.
The speech-enabled interface for AccessLine's One Person One Number(r) service represents a new frontier in user interaction with an automated personal assistant service. Built using standard speech recognition hardware and software, AccessLine complements the DTMF keypad by providing a more natural user interface to access and control the one-number, follow-me and other services.
Peter Gantchev is a Sr. Product Marketing Engineer for Dialogic Corp., responsible for Antares DSP platform and Speech Technologies.