Keyword-spotting using SRI's DECIPHER large-vocabulary speech-recognition system

The application of the speaker-independent large-vocabulary CSR (continuous speech recognition) system DECIPHER to the keyword-spotting task is described. A transcription is generated for the incoming spontaneous speech by using a CSR system, and any keywords that occur in the transcription are hypothesized. It is shown that the use of improved models of nonkeyword speech with a CSR system can yield significantly improved keyword spotting performance. The algorithm for computing the score of a keyword combines information from acoustics, language, and duration. One key limitation of this approach is that keywords are only hypothesized if they are included in the Viterbi backtrace. This does not allow the system builder to operate effectively at high false alarm levels if desired. Other algorithms are being considered for hypothesizing good score keywords that are on high scoring paths. An algorithm for smoothing language model probabilities was also introduced. This algorithm combines small task-specific language model training data with large task-independent language training data, and provided a 14% reduction in test set perplexity.<
>