ScanSoft Rekindles Dragon's Flame
It has been over five years since I last wrote on the subject of speech recognition for BYTE. My interest in the subject hasn't waned; there just wasn't much worth reporting. When I wrote my original series of articles, there were three major players in the speech recognition arena: Dragon Systems, IBM, and Lernout & Hauspie. Between 1999 and 2000, The Belgium based L&H undertook a series of acquisitions designed to make their company a one-stop shop for dictation, transcription, and machine translation services.
The acquisitions, made with L&H stock-and little or no cash-included the medical transcription assets of Rodeer Systems, Dragon Systems, and Dictaphone. The stock became the darling of Wall Street.until it became clear that the company had thoroughly cooked the books. After a messy bankruptcy, in which European banks lost $430 million in loans, Dictaphone emerged as a privately-held concern, and Dragon NaturallySpeaking became part of ScanSoft's portfolio.
The fiasco set the progress of consumer speech recognition software back not by months, but by years. As an L&H product, NaturallySpeaking incorporated only two useful features from L&H Voice Xpress-the "Nothing But Speech" filter, and improved text-to-speech capabilities. In actual use, "Nothing But Speech" worked no better than the work-around I suggested in April of 2000 for making Dragon reject random environmental noises. While text-to-speech capabilities can be helpful for catching errors in dictated documents (because spell check is of limited use when entirely wrong but correctly spelled words have been inserted), most users would choose improved accuracy over text-to-speech, given the choice.
The first signs of a turnaround came with ScanSoft's release of NaturallySpeaking 7 last year. While the main achievement of this release was the removal of bugs introduced during the Lernout & Hauspie debacle, Dragon 7 had one tantalizing feature that bode well for the future: It included the option to collect acoustic and corrections data that could be shared with ScanSoft to help improve accuracy.