The BBN BYBLOS Continuous Speech Recognition System

The BBN BYBLOS Continuous Speech Recognition System Richard Schwartz Chris Barry, Yen-Lu Chow, Alan Deft, Ming-Whei Feng, Owen Kimball, Francis Kubala, John Makhoul, Jeffrey Vandegrift BBN Systems and Technologies 10 Moulton Street Cambridge, MA 02138 Abstract In this paper we describe the algorithms used in the BBN BYBLOS Continuous Speech Recognition system. The BYBLOS system uses context-dependent hidden Markov models of phonemes to provide a robust model of phonetic coarticulation. We provide an update of the ongoing research aimed at improving the recognition accuracy. In the first experiment we confirm the large improvement in accuracy that can be derived by using spectral derivative parameters in the recognition. In particular, the word error rate is reduced by a factor of two. Currently the system achieves a word error rate of 2.9% when tested on the speaker-dependent part of the standard 1000-Word DARPA Resource Management Database using the Word-Pair grammar supplied with the database. When no grammar was used, the error rate is 15.3%. Finally, we present a method for smoothing the discrete densities on the states of the HMM, which is intended to alleviate the problem of insufficient training for detailed phonetic models. i. Introduction At BBN we have been involved in the development of Spoken Language Systems for almost two decades. As part of DARPA's Speech Understanding Research Program from 1971-1976, we developed a system that integrated continuous speech recognition with natural language understanding in a 1000-word travel management task; we call the system HWIM (Hear What I Mean). As part of another DARPA program, we have been working since 1982 on a more advanced speech recognition system based on using Hidden Markov Models. The result of this work is the BYBLOS Continuous Speech Recognition System. The basic algorithms used in the BYBLOS Continuous Speech Recognition system have been described in several papers [1, 2, 3]. In Section 2 we give a brief review of the techniques currently used in the BYBLOS system. The two features that have made the largest improvements in recognition accuracy since 1982 were the use of robust context-dependent phonetic models, and the addition of derivative spectral parameters in multiple codebooks. Each of these features used separately reduces the recognition error rate by a factor of two. Taken together, they reduce the error rate by a factor of four. In Section 3 we present the latest recognition results for the BYBLOS system. In particular, we compare the recognition results with and without spectral derivative parameters. We also demonstrate, by testing the system on training data, that the recognition accuracy is likely to improve as more training data is made available. Since several similar systems have provided test results on this database it is possible to determine the benefits of particular algorithms. In particular, we compare the error rate for using discrete densities with that using continuous densities. We also compare the recognition accuracy for speaker-dependent models with that for speaker- independent models derived from a large number, of speakers. Finally, in Section 4, we present a …

Content truncated to 500 words