Our current SpeechRecognition? software based on Sphinx3 runs fine all in all. But the setup for an individual task is cumbersome and requires several steps that are not documented well enough and most of which could be automated at least partly. There should only be _one_ place to specifiy a tasks grammar and command mapping that is easy to setup. The should also be a general LanguageModel for AtHome? that doesn't need to be touched. If so, this re-creation should also be easy for the end-user (i.e. documented and scripted). Furthermore, we currently don't have the Sphinx-version used under version control within RCSoftX. RCSoftX should be kept self-contained so that SpeechRecognition can be done on the robot without any further external dependencies.

In detail:

• Grammar
  • design or choose one uniform grammar specification format (like the one used in Sphinx4 based on EBNF with command mapping)
  • create tools that generate all necessary files and formats from that file (sphinx/DualDecoder-specs, SpeechParser, ...)
  • preferably there should be some kind of hierarchical inclusion facility to plug several grammars together to receive the final task (e.g. combine general commands used in all tests and the specific commands for a particular test)
• LanguageModel
  • The DualDecoder uses a broader language model that has to be generated from the overall speech used in AtHome?-tasks. The creation of this language model needs to be scripted and documented.
• Overall Sphinx integration in RCSoftX is missing!
  • The version of Sphinx we use needs to be included in RCSoftX and a Makefile target should be provided.
  • Data needed to do any of the above need to be put to a place that is accessible!