Speech recognition using tree-structured probability density function