November 1999

What Is Speech Activated Messaging?

BY PAUL BRENNAN

Speech activated messaging is an alternative access mechanism to the same voice or multimedia mailbox that a subscriber may already access via their desktop computer or touchtone phone. While some changes are appropriate to make best use of the capabilities of speech recognition technology, generally the structure and features of the mailbox and the messaging system will remain the same as those encountered through the traditional touchtone interface. Users should be able to easily access their multimedia mailboxes through either the traditional touchtone interface, through the speech activated messaging interface, or through a desktop computer interface.

HOW DOES IT WORK?
Speech activated messaging makes use of Automatic Speech Recognition (ASR) technology. This technology is becoming more common in a number of domains, enabling users to speak their commands and data in a growing variety of applications accessed by telephone, or run on the user's own personal computer.

Overview Of ASR Technology
ASR technology is at its core a pattern-recognition algorithm, written as computer software. Different algorithms and patterns are used depending on the requirements of the application.

Some applications require very sensitive recognition of large numbers of words (for example, dictation programs). These applications are typically “speaker dependent,” meaning that the algorithm relies on stored patterns or templates that have been customized for the individual through a “training” process in which the individual user speaks a number of words for the system. The user’s actual words are then converted into the stored templates for future use.

Other applications require general recognition of fewer words (typically less than 50) and work without the user having to individually “train” the system. This is the most appropriate choice for speech activated messaging.

The ASR software runs on the computer’s central processing unit (CPU), or on special-purpose Digital Signal Processors (DSPs). The details of the various pattern recognition algorithms are complex and highly proprietary; the better the algorithm, the more effective the ASR technology. Effectiveness means cost and quality: how much computing power is required to do the job, and how well is the job done.

CPUs and DSPs
Most voice or multimedia messaging systems make use of DSPs to offload the CPU for certain kinds of tasks. DSPs can efficiently play prompts, record voice, detect touchtone signals, and send/receive faxes. Most importantly, they can run Automatic Speech Recognition software.

DSPs do the signal processing that they are optimized for, and the CPU does the general purpose computing it is designed for. This sharing of computing load between CPU and DSPs results in systems that are more efficient, more flexible, and most cost-effective.

ASR IN SPEECH ACTIVATED MESSAGING
Here is how ASR works in speech activated messaging:

First, the messaging system determines that an incoming call should be answered by a speech activated messaging application. The CPU arranges for the connection of the caller to an appropriate DSP, one that is ready to run the ASR software.

The CPU starts the speech activated messaging application, which instructs the DSP to play voice prompts to the caller instructing them to speak their commands. (The first instruction is usually to ask the caller to log in to the system.)

The DSP is used to cancel any echo that may be present from outgoing voice prompts or messages. (Traditional telephone technology mixes together the signals of both parties, and “echo” from the outgoing signals must be removed so that incoming signals can be detected and analyzed.)

Next, the DSP compares portions of the signal with the system’s speaker-independent templates for phonemes and provides scores for how well the signal portion matches one or more of the phonemes. The DSP then uses a very complex algorithm to compare the likely phoneme(s) with the expected patterns for the command words and digits in the system's vocabulary. A final “best guess” and “second best guess” for each spoken word is returned to the application software. For example, the DSP could return “five” — 75 percent likely, and “nine” — 25 percent likely.

The speech activated messaging application software then evaluates what the speech recognizer has determined to be the most likely user input, and depending on the scores and the importance of the command word or data, either accepts the input immediately or requests that the user confirm or repeat their input

WHO WILL BENEFIT?
Speech activated messaging was developed to meet the needs of a variety of users, including those who don’t have touchtone signaling available. While touchtone penetration in North America is estimated at 70–80 percent, that still means tens of millions of dial-pulse telephones; in some countries and regions, touchtone penetration is less than 20 percent. With speech activated messaging, mobile users can now retrieve, manage, and send voice messages without touching a key, making the process much safer and more convenient.

An organization will benefit from speech activated messaging if one or more of the following is true for even some of their users:

• They commute significant distances to work and would like to manage their messages from their car phones.
• They live or travel in areas where touchtone signaling isn’t available.
• They have physical limitations to mobility or dexterity.
• They would like to manage their messages from handsets with combined keypads.
• They like to use leading-edge technology.

IMPORTANT FEATURES
Selecting the right speech activated messaging product will involve a careful analysis of the features the product offers. In some cases, the underlying speech recognition technology will provide the feature; in others, it will be provided by the service built on that technology. Regardless, if the feature isn’t there, it will make the service harder to use.

The following are some of the most important user features:

No training required: The speaker-independent recognition technology can recognize words without the user teaching the system the characteristics of their own voice. This allows a user to make productive use of the system right from the very start.

Ability to interrupt: Users can speak their commands at any time, without waiting for the system to finish prompting or presenting a message. This allows both novice and expert to interact with the system at their own pace.

Natural number input: Connected-digit recognition allows users to speak numbers naturally, without artificial and time-consuming pauses between digits.

Multi-stage recognition: If a user encounters difficulty using the system, due to background noise or other factors, the system continues to perform. Confirmation of spoken commands is used when there is a likelihood of misrecognition, and in severe circumstances a single-digit numeric entry mode replaces natural number input.

Touchtone signaling, if desired: Users may use touchtone interface to log in or during any numeric entry, and if desired, can switch from speech activated messaging to touchtone messaging at any time during their session. This feature is also important from a security standpoint: it would be inappropriate to speak a user ID and password in a public place.

Multilingual support: Users speak different languages, and your speech activated messaging service should be capable of dealing with users in their native languages. Users forced to speak a second language may have more difficulty with the Automatic Speech Recognition.

IN THE FUTURE …
Speech activated messaging services will mature hand in hand with the underlying Automatic Speech Recognition technology. As accuracy increases, more messaging features will be added without impairing day-to-day activities.

When selecting a speech activated messaging product, insure that the product can grow with the ASR technology, and take advantage of the performance and accuracy improvements that can be expected in the coming years. You may also wish to determine where the ASR technology comes from: some ASR providers are still spending significantly on research and development of ASR, and it would be to your benefit to be able to make use of the advances in the technology in the future.

Paul Brennan is a product planner for multimedia messaging at Nortel Networks. Paul has done user interface design, feature architecture, and product planning for Nortel Networks for 17 years. Nortel Networks delivers the combined expertise and proven leadership in telephony and IP-based data, wireline and wireless networking to bring all your communications together over a single network. For more information, please visit their web site at www.nortel.com