Speaker adaptation by means of adjustment of speaker characteristic properties, such as vocal tract length, has the important advantage compared to conventional adaptation techniques that the adapted models are guaranteed to be realistic if the description of the properties are. One problem with this approach is that the search procedure to estimate them is computationally heavy. We address the problem by using a multi-dimensional, hierarchical tree of acoustic model sets. The leaf sets are created by transforming a conventionally trained model set using leaf-specific speaker profile vectors. The model sets of non-leaf nodes are formed by merging the models of their child nodes, using a computationally efficient algorithm. During recognition, a maximum likelihood criterion is followed to traverse the tree. Studies of one- (VTLN) and four-dimensional speaker profile vectors (VTLN, two spectral slope parameters and model variance scaling) exhibit a reduction of the computational load to a fraction compared to that of an exhaustive grid search. In recognition experiments on children’s connected digits using adult and male models, the one-dimensional tree search performed as well as the exhaustive search. Further reduction was achieved with four dimensions. The best recognition results are 0.93% and 10.2% WER in TIDIGITS and PF-Star-Sw, respectively, using adult models.
Bibliographic reference. Blomberg, Mats / Elenius, Daniel (2009): "Tree-based estimation of speaker characteristics for speech recognition", In INTERSPEECH-2009, 580-583.