Reduction techniques for frames and frame dimensions in automatic speech recognition