Lecture Browser allows you to search for keywords in videos of classes and lectures on the Internet
According to this Technology Review article, MIT researchers have developed an audio and video search tool that solves one of the main problems in this field: how to divide a large academic class into manageable chunks, pinpoint keyword location, and direct to users towards them. This search engine builds on decades of voice recognition research by MIT and other institutions to convert audio to text and enable searches within it.
The Lecture Browser has emerged at a time when more and more universities are posting videos and podcasts of their classes online. Although its content can be very useful, locating specific information within classes is difficult and sometimes frustrating for students, used to finding what they need in less than a second with Google.
"It's one of the biggest problems for universities across the country," says Jim Glass, a researcher at MIT. "It is a real challenge to know how to spread them and facilitate students' access to specific parts of the class in which they may be interested."
The fundamental elements of the Lecture Browser have been around for more than 30 years in research laboratories at MIT and elsewhere, such as BBN Technologies, Carnegie Mellon, SRI International or the University of Southern California. Their initiatives have resulted in software that is ultimately good enough to guide the average person, says Premkumar Natarajan, a scientist at BBN.
A few companies, such as online video and audio search engines Blinkx and EveryZing, are already using software that converts audio into searchable text, but the MIT researchers ran into some specific problems using it for lectures. academic. For one thing, English is not the mother tongue of many of the speakers, making it difficult for automated transcription systems trained to work with American accents. On the other hand, the words that predominate in science classes can often be little known. Finally, says Regina Barzilay, a computer science professor at MIT, classes tend to have a very poorly discernible structure, making it difficult to divide and organize their content to facilitate searching.
To solve these problems, the researchers first set up the software that converts audio to text. They trained the software to understand certain accents using precise transcriptions of short chunks of audio recordings. To help the software identify unusual words, the researchers provided it with additional data, such as text from books and lectures, that help the software accurately transcribe four out of five words. However, if the system is used with a person whose mother tongue is not English and for whose accent and vocabulary the system has not been trained, the precision may drop to 50% (such a low precision would not be useful for a transcription, but yes for a keyword search).