Cool Conversations LLCs first release of its Cool C Talking Computer software is a good demonstration of how far speech recognition has comeand how far it still has to go. Cool C is one of the first applications to use Microsoft Corp.s Microsoft Speech Engine to bring speech recognition and interaction capabilities to the user.Cool C requires an Intel Corp. Pentium-based PC, 128MB of RAM, 110MB of free disk space, Windows NT or later, and speakers and a microphone. The software also requires Microsoft Speech Engine, a feature in Windows XP. (The engine is included with the Cool C installation disk for users of earlier operating systems.) eWEEK Labs tested Cool C Talking Computer (priced at $19 per user), which uses speech technologies to verbally respond to spoken questions, and the Cool C Read Write module ($10 per user), which converts spoken word to text. Cool C Talking Computer shipped in February. The Read Write module became available in March. Installing the Cool C Talking Computer application and the Read Write module was quick and easy, but training Microsoft Speech Engine to recognize voice was a different story. When Cool C is installed, users are prompted to go to a Voice Training Wizard within Windows. To train the engine, we read text for at least 30 minutesoften reading the same phrases repeatedlyso the software could learn the nuances of our voice. Microsoft has an initial database of sounds, but voice training is essential to Cool Cs ability to recognize spoken words. As long and arduous as the training period is, it must be noted that Cool C probably cannot reach maximum accuracy until after a week or more of regular use. Rival products need a similar amount of time to get up to speed. To test the softwares ability to recognize and translate speech into text, we read the first paragraph of Norton Justers "The Dot and the Line: A Romance in Lower Mathematics" to the Cool C Read Write module using an LVA-7330 headset microphone from Labtec Inc. When we read the test paragraph, "Once upon a time, there was a sensible straight line who was hopelessly in love," Cool C initially translated it as "Unsupported on the DOS and sensible a state line test closely in lot." Cool Cs ability to recognize words and phrases improved as we continued to spend more time training Microsoft Speech Engine. Although we never came close to 100 percent accuracy, we were able to get at least 65 percent accuracy after two days of testing. In general, more time spent training Microsoft Speech Engine translates into higher accuracy. However, demanding professionals who do a lot of dictation would be better off looking for productivity gains from an enterprise-class product such as Scansoft Inc.s Dragon NaturallySpeaking. As with most speech recognition applications, Cool C can quite capably read text back to a user. The application had no problem reading text from Web pages, e-mail messages or Word documents in tests. The applications integration with Microsoft Speech Engine was particularly apparent during interaction tests with the application. We were able to carry on conversations with our test clientas limited as those discussions may have beenon the weather, the date, the time and the U.S. Capitol. When we asked Cool C a question it could not decipher, the program would tell a joke to keep our conversation going. The tools conversation database can be expanded to offer customized replies when asked a question. Senior Writer Anne Chen can be reached at firstname.lastname@example.org.
Click here to read eWEEK Labs review of Microsoft Speech Server 2004 beta.