A Speech to Speech Translation System Built From Standard Components

A SPEECH TO SPEECH TRANSLATION SYSTEM BUILT FROM STANDARD COMPONENTS Manny Rayner 1, Hiyan Alshawi 1, Ivan Bretan 3, David Carter a, Vassilios Digalakis 2, BjSrn Gambiick 3, Jaan Kaja 4, Jussi Karlgren z, Bertil Lyber9 4, Steve Pulman 1, Patti Price 2 and G-'hrister Samuelsson 3 (1) SRI International, Cambridge, UK (2) SRI International, Menlo Park, CA (3) SICS, Stockholm, Sweden (4) Telia Research AB, Haninge, Sweden ABSTRACT This paper I describes a speech to speech translation system using standard components and a suite of generalizable cus- tomization techniques. The system currently translates air travel planning queries from English to Swedish. The modu- lax architecture is designed to be easy to port to new domains and languages, and consists of a pipelined series of process- ing phases. The output of each phase consists of multiple hypotheses; statistical preference mechanisms, the data for which is derived from automatic processing of domain cor- pora, are used between each pair of phases to filter hypothe- ses. Linguistic knowledge is represented throughout the sys- tem in declarative form. We summarize the architectures of the component systems and the interfaces between them, and present initial performance results. 1. INTRODUCTION From standard components and a suite of generalizable customization techniques, we have developed an English to Swedish speech translation system in the air travel planning (ATIS) domain. The modular architecture con- sists of a pipelined series of processing phases that each output multiple hypotheses filtered by statistical pref- erence mechanisms. 2 The statistical information used in the system is derived from automatic processing of domain corpora. The architecture provides greater ro- bustness than a 1-best approach, and yet is more com- putationally tractable and more portable to new lan- guages and domains than a tight integration, because of the modularity of the components: speech recognition, source language processing, source to target language transfer, target language processing, and speech synthe- sis. Some aspects of adaptation to the domain task were fairly simple: addition of new lexical entries was facil- itated by existing tools, and grammar coverage required 1 The research reported in this paper was sponsored by Swedish Telecom (Televerket Ngt). Several people not listed as co-authors have also made contributions to the project: among these we would particularly like to mention Marie-Susanne AgnKs, George Chen, Dick Crouch, Bsrbro Ekholm, Arnold Smith, Tomas Svensson and TorbjSm ~hs. 2The preference mechanism between target language text out- put and speech synthesis has not yet been implemented. adding only a few very domain-specific phrase structure rules, as described in Section 3.1. Much of the effort in the project, however, has focussed on the development of well-specified methods for adapting and customizing other aspects of the existing modules, and on tools for guiding the process. In addition to the initial results (Section 5), the reported work makes several contribu- tions to speech translation in particular and to language processing in general: A general method for training statistical preferences to filter multiple hypotheses, for use in ranking both analysis and translation …

Content truncated to 500 words