HOW TO RECOGNIZE SPEECH? OR HOW TO WRECK A NICE BEACH....

HOW TO RECOGNIZE SPEECH? OR HOW TO WRECK A NICE BEACH....
1The machine isn't really thinking. It's slicing Jane's words into tiny segments and analyzing the resonant frequencies. The results are matched against "templates"--idealized pieces of real words. Next, the computer calculates the probability that certain clusters of sounds correspond to syllables or words. Many syllable sequences are possible, such as "re-cog-nize" and "wreak a nice." The software chooses by dicing sounds into two- and three-word strings, and ruling out unlikely alternatives.	2 Having correctly identified most of the words, the software dumps those that don't matter, such as "talk about." Now, it interprets Jane's meaning, drawing on stored knowledge and statistical analysis. In the dialogue above, the computer runs into two syntactically valid alternatives, so it queries Jane. Computers today do not handle such situations well. And Jane makes life even harder by interrupting in mid-sentence, or "barging in." Eventually, however, computers will master dialogue.	3 Having confirmed Jane's wishes, the machine starts its hardest task: finding information in the sprawling database called the Web. It translates Jane's request into a database query, just as today's search engines do. It may even be smart enough to extract a synopsis from the first few paragraphs of a long written statement in an annual report. Someday, computers will be able to deftly summarize articles by parsing sentences into subjects, verbs, and objects so they can grasp events.	4 Summarization is just the beginning. Scientists are exploring ways to tag data that are posted on the Web, so computers can quickly find what they are looking for--even data in a pie chart or bar chart--and present it in spoken sentences. Today, Jane's computer could tell her that Microsoft has posted a video clip about speech recognition on its Web site. It couldn't describe the video images. But it could convert the accompanying audio track to text and then summarize it.

Return to main story