A User Interface-Level Integration Method for Multiple Automatic Speech Translation Systems

A User Interface-Level Integration Method for Multiple Automatic Speech Translation Systems Seiya Osada1, Kiyoshi Yamabana1, Ken Hanazawa1, Akitoshi Okumura1 1 Media and Information Research Laboratories NEC Corporation 1753, Shimonumabe, Nakahara-Ku, Kawasaki, Kanagawa 211-8666, Japan {s-osada@cd, k-yamabana@ct, k-hanazawa@cq, a-okumura@bx}.jp.nec.com Abstract. We propose a new method to integrate multiple speech translation systems based on user interface-level integration. Users can select the result of free-sentence speech translation or that of registered sentence translation without being conscious of the configuration of the automatic speech translation system. We implemented this method on a portable device. Keywords: speech translation, user interface, machine translation, registered sentence retrieval, speech recognition 1 Introduction There have been many researches on speech-to-speech translation systems, such as NEC speech translation system[1], ATR-MATRIX[2] and Verbmobil[3]. These speech-to-speech translation systems include at least three components: speech recognition, machine translation, and speech synthesis. However, in practice, each component does not always output the correct result for various inputs. In actual use of a speech-to-speech translation system with a display, the speaker using the system can examine the result of speech recognition on the display. Accordingly, when the recognition result is inappropriate, the speaker can correct errors by speaking again to the system. Similarly, when the result of speech synthesis is not correct, the listener using the system can examine the source sentence of speech synthesis on the display. On the other hand, the feature of machine translation is different from that of speech recognition or speech synthesis, because neither the speaker nor the listener using the system can confirm the result of machine translation. Thus, an error in the machine translation is critical in a speech-to-speech translation system. Instead of machine translation, there is a parallel text-based translation which uses parallel bilingual sentences registered in the system. This retrieves the corresponding translation by referring to the registered source sentence. However, in parallel text-based translation, although the quality of translation is largely guaranteed, only a limited number of sentences can be translated, because it is impossible to cover all the utterances of users by the parallel text of source language. When no registered sentences correspond to what the user says, he or she has to choose a most preferable one among retrieved sentences that roughly reflect the utterance; otherwise the speaker may give up the attempt. Accordingly, an integrated method which is easier to use is required in which the accuracy of translation becomes compatible with coverage by integrating these two translation components. This paper proposes a new method to unify the automatic speech translation system with free-style sentence translation component and that with parallel text-based translation component at the user interface-level in order to further ease the operation of users compared to the previous unified approach. 72 2 Previous approach: Method-level integration To solve these problems on the machine translation and the parallel text-based translation, there is a system which integrates two components -- for free-sentence translation and for parallel text-based translation [4]. This system (see Figure 1) has a speech recognition component, a speech …

Content truncated to 500 words