Raytheon BBN Technologies

Automatically finding information in speech, such as topics, events and named entities is one of the main goals of the speech recognition process. At Raytheon BBN Technologies we are developing techniques that can accommodate various levels of acoustic resources that are available to a user. Even in cases where there are no transcriptions available or where there may not be a written form for the language under consideration, we have algorithms that can provide useful information for a user.

In the case of very limited knowledge about the sounds of a spoken language, a user need only provide the information extraction system with examples of speech that are on a topic or event of interest. Based on this input from the user, a speech recognizer and topic/event classifier are automatically trained to extract from the audio data passages that are on the desired topic or event. The automatic training process of the speech recognizer does not require any input about the sounds of the language from the user. The automatic training algorithms will discover, from the speech data, a set of phone and syllable like units to use in the recognition process.