Summary: parameters are updated using GEM. From the table, we
notice a modest improvement from adding the nonlinear
In this paper we have presented a novel nonlinear transfor
mation based adaptation algorithm. The nonlinearity was
implemented by a multilayer perceptron, whose weights
were estimated using a generalized EM algorithm. This
technique made use of gradient ascent training embedded
within each maximization step of the EM algorithm.
Experimental results were presented for nonnative speaker
adaptation. So far, our results show a modest improvement
using the nonlinear technique as compared to previous lin
We are continuing to explore this technique to see if we
can further improve our performance. We also intend to
investigate the application of nonlinear adaptation of
HMM parameters for other types of acoustic mismatches,
such as noisy speech.