Attelage de Système Hétérogènes – ASH

Project ASH proposes a new approach for combining several Automatic Speech Recognition (ASR) systems. This approach consists in making ASR systems exchange information during the decoding process, on the fly, whereas the classical approach in this field consists in only combining final outputs of ASR systems. This implies the design of a specific framework, which allows for the use of various recognition techniques: landmark-driven decoding will be particularly focused on for integration within this framework. For the approach developed in project ASH, we intend to take advantage of the current trend toward multi-core CPU architectures by parallelizing the execution of several ASR systems processing the same audio data. This should yield a more robust system with better recognition results while keeping processing time as low as possible ' possibly near real time. The three partners involved in the project have already collaborated on preliminary works concerning system combination: the preliminary results, which were validated by the international scientific community, were very promising and confirmed the partners in wanting to develop this project.