Many of the so-called future capabilities of speech recognition cited in Janelle Brown's article are actually available now. Such things as using voice to turn on and off lights and other appliances in a house, or to surf the Internet and download an article, are commonplace to many people. Perhaps that is because speech recognition technology has been in development since the 1950s, not the 1970s.

Ms. Brown's idea that speech recognition software can "stunt the creative writing process" may be true for her, but it is certainly not true for me or for many, many other users. In fact, I find that my communication is much more clear when I use Dragon NaturallySpeaking than when I hunch over the keyboard and type a document. I find speech recognition stimulating and interesting, and it seems as if the documents I create using it are better documents because they are somewhat more conversational.

Finally, there is the matter of chatter in the workplace. Ms. Brown seems to think everyone will need to wear earplugs, if everyone is using speech recognition in the office. As a matter of fact, everyone is chattering now in the office, using another instrument of technology called the telephone. Speech recognition need be no different. I doubt that we will get rid of telephones for this reason!

-- Van Thompson

As the author of a manual for a voice recognition program developed for BellSouth Cellular Corp.'s Research and Development Lab, I felt that the author failed to provide some important distinctions between types of voice recognition programs.

The systems that have limited commands, such as systems that navigate voice mail menus, are created by collecting samples from a large number of speakers and then developing templates for each word that can accommodate wide variations in articulation of the words. These systems can be speaker-independent because they use huge samples as the base of the system. Limited vocabulary systems can recognize voices with a high degree of accuracy, but collecting all the samples and creating templates is too time consuming for the development of large vocabularies. In addition, a system based on collected samples does not handle the natural joining of words that occurs in continuous speech.

Unlike Dragon Naturally Speaking and ViaVoice, limited-vocabulary systems already function accurately enough for most people to use them without too much frustration, as long as the user knows how to give the commands to the system (the commands must be given slowly and in the right order). I suspect that the AutoPC program discussed at the end of the article is a limited-vocabulary system.

The Dragon Naturally Speaking and ViaVoice programs which are the focus of the article, however, are large-vocabulary systems. Large-vocabulary systems are built by collecting phoneme patterns and stringing those together. Phonemes are the sounds that make up the English language: ah, ae, etc. There are only about 70 phonemes, so it's easy to see why a large-vocabulary system must be built out of these building blocks instead of templates of individual words. All of these phoneme-based systems require training to a particular user. Developing large vocabulary systems involves a level of complexity that developing a limited vocabulary system does not. Because of the additional complexities involved in recognizing natural speech, large-vocabulary systems are not yet capable of recognizing speech with the same accuracy as the limited-vocabulary systems.

-- Suzanne LaForest

Next page | The trouble with "Trek"