Project 9: Hybrid episodic-abstract computational modelling: ASR
Two fundamentals of traditional HMM ASR systems, namely generalization of phoneme-like units and a top-down search procedure, are in conflict with current understanding of HSP, i.e. the preservation of FPD and the importance of bottom-up processing. Properties of recognition of novel words and L2 learning require early abstraction of (sub)-phoneme-like units; while other phenomena hint at detailed segmental matching. Project 9 seeks to further improve its recent combined episodic and abstractionist SR model. The focus for the first half of the project will be on weighting acoustic-phonetic information relative to more abstract knowledge. Later this project may seek to incorporate systematic linguistic knowledge fed to it by Project 10, using the bottom-up information in ways that humans are thought to.
Working on this project: » Prof Dirk Van Compernolle » Prof Roger Moore » Dr Sébastien Demange » Dr Kris Demuynck » Dr Dino Seppi
