Toward a Real-Time Spoken Language System Using Commercial Hardware

Toward a Real-Time Spoken Language System Using Commercial Hardware Steve Austin, Pat Peterson, Paul Placeway, Richard Schwartz, Jeff Vandergrift BBN Systems and Technologies Inc. 10 Moulton St. Cambridge, MA, 02138 Abstract We describe the methods and hardware that we are using to produce a real-time demonstration of an integrated Spo- ken Language System. We describe algorithms that greatly reduce the computation needed to compute the N-Best sen- tence hypotheses. To avoid grammar coverage problems we use a fully-connected first-order statistical class grammar. The speech-search algorithm is implemented on a board with a single Intel i860 chip, which provides a factor of 5 speedup over a SUN 4 for straight C code. The board plugs directly into the VME bus of the SUN4, which controls the system and contains the natural language system and application back end. 1. Introduction One goal of the Spoken Language System (SLS) project is to demonstrate a real-time interactive system that inte- grates speech recognition and natural language processing. We believe that, currently, the most practical and efficient way to integrate speech recognition with natural language is using the N-Best paradigm, in which we find the most likely whole-sentence hypotheses using speech recognition and a simple language model, and then filter and reorder these hypotheses with natural language. Although we claim that this process is efficient, finding the N-Best sentences still requires significant amounts of computation. To accelerate the speech recognition process, several sites have been developing special-purpose hardware, from fully- custom VLSI [5] to custom boards based on general-purpose processors operating in parallel [1]. However, with the rapid changes in commercial hardware in the past few years we felt that if we could achieve real time with commercially available hardware we would realize a number of advan- tages, such as decreased development time, increased relia- bility, and better hardware and software support. We would not have to divert attention from speech-recognition research and our results wouM be more easily shared with the commu- nity. We felt that these hopes were not unrealistic because, for our algorithms in particular, which are research-oriented and seldom have a stable, regular StlUCture, the potential gains from custom hardware wouM not be that great. Our strategy in developing a real-time spoken language system has been to find appropriate commercial hardware and then tailor our algorithms to the hardware and applica- tion. In this context, we are trying to approach the following goals: 1. real-time processing with a short delay 2. reasonable-cost, commercially-available hardware 3. source code compatibility with our research programs 4. computation of the N-Best sentences for large N 5. use of a robust fully-connected statistical grammar 6. practical memory requirements 7. negligible loss of accuracy 8. ability to handle large vocabulary Two of the new algorithms used in this effort are described in a separate paper [8]. Specifically, the Word-Dependent N- Best search and the Forward-Backward Search. In Section 2 we describe the class of hardware that we have chosen for this task. Section 3 reviews the reasons …

Content truncated to 500 words