Computer science graduate Stephen Piron dreams to revolutionize the surfing of media files in the same way that Google revolutionized websurfing. Last month, his startup, SimonSays Voice Technologies, unveiled software that transcribes audio or video content.
Currently, major search engines, such as Yahoo and Google, merely parse through a file’s so-called metatags. These are terse descriptions of the file, input by the user. Another technique is looking at the text adjacent to media on the webpage, which often offers clues about its content.
But while such searches would succeed in retrieving a film documenting a canoe trip through Algonquin Provincial Park, say, you would not be able to search for and skip to the part of the film where the protagonists encounter moose, or other wildlife.
SimonSays takes a different tack. “It’s a server-based solution,” said Piron, “so you would log on and upload your audio, and it would come back to you as text.”
This allows SimonSearch to throw more computing power at it. But their software also works differently than packages available for personal computers, for which the user must first read some passages into a microphone.
“We’re … training a language model, rather than a voice profile,” Piron explained. Language models work out the probabilities of certain chains of words preceding each other.
Their proof of concept is online, at radiophiles.org, a database of transcripts of interviews conducted by Jennifer Leonard, host of the CIUT show Dialogue. SimonSays also transcribes scripts for movie studios. Medical transcrption might be another untapped market.
And scribblers? “I didn’t really think about journalists,” Piron said, adding that their software might struggle with poor quality recordings.
SimonSays now aims to hitch its transcription prowess to the search power of Google or Yahoo. “That would be a perfect fit,” Piron remarked.
