next up previous contents
Next: Experiments Up: Task 4.1: Technical Description Previous: Phone to phone transition

Discussion

Such a model is expected to lead to better recognition performance in the case of training independent tasks. Indeed, the training data always contains the phonemic units in a limited number of left and right contexts. In standard recognisers (and standard reference tasks), it is possible to make use of this contextual information to improve performance, e.g., by using context dependent phone model. Even when using context independent phone models (which is often the case with hybrid HMM/ANN systems that are however known to yield comparable - although still somewhat lower - performance compared to context-dependent models), the phone models will implicitly capture some contextual information. However, if the application (or test) vocabulary is different from the training vocabulary, using that phonemic contextual information could result in a loss of performance since the trained models are no longer really appropriate. The context dependent hybrid system attempts to limit this effect by tying the distributions of context dependent models. In this way, it can be expected that all major contextual effects will be captured by those tied ``transition'' states while the context independent phoneme models will focus on the actual ``steady-state'' section of each phonemic segment.


next up previous contents
Next: Experiments Up: Task 4.1: Technical Description Previous: Phone to phone transition
Christophe Ris
1998-11-10