next up previous contents
Next: Bibliography Up: WP5: NEW ARCHITECTURES AND Previous: Task 5.5: Technical Description

WP5: Conclusion

The fundamental assumptions leading to multi-band ASR have been investigated. Multi-band speech recognition has been generalized to a ``multi-stream'' approach. A first step to the full REMAP training has been taken by the implementation of a forward-backward training for MLPs. Syllable based training has been tested, and input feature space dimensionality reduction has been investigated. Context dependent phone modelling has been tested for recurrent networks.



Christophe Ris
1998-11-10