オステンドーフ マリ,シンガー ハラルド
Modification of SSS for
Speaker-Independent HM-Net Design
Abstract:This report describes a new approach to successive state splitting that uses the maximum likelihood criterion directly in choosing specific splits, rather than first choosing the state with the most variability and then choosing the best split for that state. By testing splits directly, the algorithm is better suited to speaker-independent HM-Net training, which we believe will improve recognition performance on spontaneous speech. Preliminary experiments are described for assessing recognition performance gains and training cost reduction associated with the algorithm.