Next: Task 5.2: Global Discrimination
Up: Task 5.1: Perceptual Models
Previous: Combining multiple time-scale features
It appears that sub-band alignments can have significant timing
deviations from the full-band alignments; thus, we would expect that
there is a potential for improvements in acoustic modeling if longer
time-scale information stream merging (i.e., phone or syllable) is
Preliminary work on multiple stream recognition
will now be extended in several directions,
Long-term features: Further experiments have to be done
to determine the features that are best suited
to capture long-term dynamic properties.
Recombination criterion: So far, only a likelihood based
recombination has been tested.
Further work with subband ASR.
Combining subband and multiple temporal scale recognition.