By now, everyone is familiar with voice mail technology. You call a business and a voice directs you to use your touch-tone phone to direct your call or to leave a voice message. Of course if you don't have a touch-tone phone, the current voice mail technology isn't accessible, and you need to wait for the operator to help you - unless the system contains a neural network.
By using BrainMaker to train neural networks for speech recognition, Dr. Mark Ortner of Compass Technology in Sarasota, Florida has developed P.C. based software that will make voice recognition technology affordable to small and medium sized businesses around the world. Soon, you'll be able to reach the voice mail box of the desired party by phone or fax whether you have touch-tone phone or not.
In 1992, Compass Technology was acquired by Octel Communications, the world's largest provider of voice and fax information processing and services. Currently, Dr. Mark Ortner is revolutionizing information processing. So far, Dr. Ortner has trained a network using 2500 facts and 28 words, including the numbers "zero" through "nine", the words, "yes" and "no", and the names of various departments within the company. The degree of recognition accuracy ranges from 90-97%.
The current application is for voice recognition of an extension. Dr. Ortner collects "voice data" by having a variety of callers dial an extension and enter a "mailbox" located at the "voice training center." The callers' words are run through a normalizer (confidential in nature) which creates a BrainMaker output file (fact file). That information is then trained on a neural net.
According to Dr. Ortner, the advantages of this kind of system are far-reaching. In the United States, only about 37% of the telephones are not touch-tone. However in most other countries, touch-tone is nonexistent. By providing a trainable voice recognition system, the purchaser could record the words used at their location, train the network, then run it. A caller could say, "Extension 230" and the system would make the correct phone transfer. Touch-tone is no longer necessary.
Dr. Ortner's network has 400 input neurons, 107 hidden neurons and, at the present time, the output layer has 28 neurons. This will eventually change as more words are added. The output of the network is the recognition of the spoken word which is drawn from a symbol table.
Dr. Ortner uses BrainMaker for training the neural net, but wrote the software that actually runs the neural net. (BrainMaker Professional comes with this code as part of the Runtime License). Since the largest system Compass Technology handles is 32 ports, this kind of voice recognition system is ideal for small and medium sized businesses.
Once this system is on the market, Dr. Ortner will turn his attention to developing a phonic-based system. The caller will be able to pronounce a word and have the system convert it directly to text. This would be a big advantage to the deaf. A deaf person could read a voice message as it is printed on screen or print a message back and have it translated into voice.