Hidden dynamic models of speech co-articulation
 

Research Issues for WS98

 
 
 
Entities
Relationships
Comments, issues
words
phonemes
phonology "pronumciation modeling"
phonetic elements allophones?
alignments
targets one (vector) target per phonetic element type.  dimensionality?
dynamics time constants, shapes, overlap, dependence on phonetic label
dynamic state Formants? Vocal Tract Resonances? vocal tract shape? ...
mapping Form, invertability, learning, algorithms, scalability, initialization
acoustic pattern mfcc, ...
acoustic metric Euclidean distance in transformed log spectrum,  (perceptually motivated?)
data artificial, synthetic speech, selected,  laboratory, field,  speakers.
testing/use acoustics -> dynamic state -> HMM
score aligned transcripts, select
synthesize unseen contexts
score lattice?
new recognizer??