Conversational Speech Transcription and Retrieval System

Accurate speech recognition of informal, interrupted, colloquial conversational speech is fundamentally a difficult task. To compensate for this, the C-STARS speech recognizer does not just produce a single best transcript, it also produces alternative hypotheses, along with confidences for each word.

C-STARS uses a powerful word-based speech recognizer for maximum accuracy for most of the words in a language. In addition, C-STARS generates phonetic information that supports unlimited vocabulary keyword search, including new and uncommon terms, such as rare names of people and places. Keyword queries are entered in the original language with no need to guess the phonetic spelling. This hybrid approach leverages both word and phonetic speech recognition to provide high accuracy with no restrictions on the words that can be searched.

C-STARS displays the automatically generated transcript of each audio recording that contains a possible match. Although the automatic transcription contains errors, an analyst can quickly read a keyword match and the surrounding transcript to determine if the result is of interest. If there are questions about what was said, the user can click on any word in the transcript to hear the corresponding speech.

Behind the scenes analytics let C-STARS users quickly find and focus on unusual or interesting terms that should be prioritized for more careful listening or transcription. Word clouds make it easy for users to find the most salient words in a collection. C-STARS also detects and highlights topics in a conversation.

C-STARS is fast - searching thousands of hours of audio per second so users can try different queries to quickly get a sense of the characteristics of an audio collection. As a pioneer in speech and language technologies and a contractor on some of the U.S. government’s largest R&D programs, BBN has a long track record of excellence. Our core technology is established, robust, and state-of-the-art.