Tutorial Announcement      

Coling 2000 Tutorial Announcement

Andreas Klüter and Norbert Reithinger,
Verbmobil: the development and integration of a large
speech-to-speech translations system

Verbmobil is one of the largest speech-to-speech translation systems developed in the last decade. It was developed at more than 30 institutions in Germany and the USA. The design and development of a system in such a heterogeneous environment posed challenges in various dimensions. The goal of this tutorial is to give the participants a feeling of the functionalities of a large system. We also address the various technical problems and their solution and stress the important of the human factors in such a development process.

We will first give an overview on the partners, the general approach, and the tasks we had to solve. In the main section we present a guided tour through the system, both with slides and demonstrations using the Verbmobil prototype system from the audio input in the source language to the synthesis in the target language.

We start with recording, audio coordination, and speech recognition. The output of the recognition are word lattices for each turn. Prosodic information about boundaries and stressed words is added. The four translations tracks - example based, dialogue act based, statistical, and deep translation consisting of syntactic and semantic processing, transfer and generation - consume the lattice. We will also look at data exchanged between the modules and introduce the VIT formalism. Before synthesizing the translation, a special module selects the best translations from the four translation tracks.

The grounding for a successful system is a sound software base. The testbed developed in Verbmobil is a major success story, since it enables distributed development and stable, indirect communication between the functional components. It also includes a rich variety of test and debugging tools we will present.

A very important aspect are the non-technical issues in a large project with lots of partner sites from different areas of speech technology. We show how the partners were coordinated and which management structures were important for the development of the system.

related events
  DFKI Language Technology Lab
German Research Center
for Artificial Intelligence
Language Technology Lab