A path-stack algorithm for optimizing dynamic regimes in a statistical hidden dynamic model of speech