Language Barriers along the Long Tail of Language
Event details
| Date | 18.05.2009 |
| Hour | 16:15 |
| Speaker | Prof. Alex Waibel, Carnegie Mellon University, USA & University of Karlsruhe, Germany |
| Location |
INM202
|
| Category | Conferences - Seminars |
During the last decade Speech-to-Speech Translation has morphed from an exotic subfield at speech & language conferences to the main and best funded research topic in the speech and language area. During its rapid rise the field has gone through distinct phases:
1) domain limited concept demonstrations
2) domain limited spontaneous two-way translators
3) domain-unlimited speech translators.
The rapid growth is explained by considerable commercial, public and humanitarian interest and the intense R&D activity has led to considerable progress and impressive new capabilities. However, we are still far from having overcome the language divide: Aside from continuing issues with performance, most advances have also only been achieved on a handful of well-funded languages. To reach the goal of global connectedness this trick would have to be repeated for many more (6,000+) languages, not to mention dialects and accents. With present methods, the cost would be prohibitive and so we see the emergence of the latest challenge, (in our view the 4th phase): how to go from 10 to 1000+ languages.
At first glance, the problem may just be a repetition of current methods, perhaps with economies of scale. On closer inspection, however, it becomes clear that language data acquisition & maintenance represent the most significant cost, and they are inversely proportional to the size of a language. They can be addressed by advancing machine learning and human interaction techniques that lead to lowered cost and more dynamic language maintenance. Available solutions fall into five orthogonal categories that will be discussed.
Alex Waibel's bio
Practical information
- General public
- Free