DSTO in Voice Software Breakthrough

The federal government's Defence Science and Technology Organisation (DSTO) has teamed with a small Melbourne software company to crack one of the big challenges in voice recognition software: automatic transcription software that lets multiple people speak at once.
The new application, demonstrated at the CeBIT Australia technology expo in Sydney yesterday, creates transcripts of both sides of a two-person conversation, according to Greg Findlay, in charge f global market development at Voice Perfect Systems, a privately held software company that worked with DSTO on the six-year project to create the software.
The application, he said, also identified who was speaking and figured out what they were saying to a high degree of accuracy. It was initially developed by DSTO to automatically create written records of defence command and control operations.
Mr Findlay said the commercial version, known as Voice Perfect Multi-Speak, would eventually be capable of transcribing meetings of more than two people, making it useful for automatically taking minutes of boardroom meetings. But the system has a catch that could make the more image conscious board member think twice before installing it: every voice requires its own microphone, and in demonstrations of the software yesterday, that involved each party wearing a call-centre-style microphone headset. Mr Findlay said the company was working on several ideas that could allow Multi-Speak to work without separate microphones for each participant.
One way would be to use tabletop microphones that electronically zoomed in on whoever was speaking. But that technology, which is already commercially available, did not pass on to Multi-Speak information about who it was zooming in on.
So it might be that people in meetings would have to wear lapel mikes, Mr Findlay said. The new software is based on the popular speech recognition engine, Dragon NaturallySpeaking, made by ScanSoft.