The goal of the Neotec Smarter Interpreting project was to provide interpreters with the possibility to benefit from suggestions obtained from direct ST models during simultaneous interpretation sessions. Toward this goal, we developed and evaluated a number of novel solutions that allowed us to get significant improvements over the state-of-the-art systems existing at the beginning of the project. Our activities can be summarized as follows:
- We investigated the weaknesses of direct ST models with respect to the NEs and domain-specific terminology, proposing techniques to increase their accuracy in rendering person names;
- We introduced the first direct ST models capable to jointly translate and recognize NEs from the input speech;
- We evaluated different methodologies to perform real-time translation with direct ST models;
- We combined all these aspects into a single system, which was exposed through a WebSocket interface and connected to the Smarterp CAI interface;
- We tested our solution in two demos with real users to assess their overall usefulness, obtaining positive feedbacks.
In light of the above contribution, the project has successfully served its purpose by leading to the first CAI solution based on direct ST systems, which displayed an effective quality-latency trade-off.
The software related to the project is released with GPL 3.0 license and is available at: https://github.com/hlt-mt/smarterp.