next up previous contents
Next: Task 4.2: Unsupervised Adaptation Up: Task 4.1: Training Independent Previous: Results

WP4 Future Development

We have proposed an new approach for incorporating phonetic context dependent modelling in the hybrid system. We now intend to compare this method to other hybrid context dependent systems such as described in [] and [5] both trying to estimate the posterior probability of a phone class qk in the context Cl given the acoustic vector x(t) :

P(qk, Cl | x(t)) = P(Cl|x(t)) * P(qk|Cl, x(t))

The difference between the two methods mainly stands in the way P(qk|Cl, x(t)) is estimated.

Christophe Ris