The Viterbi decoding algorithm is a formal technique for finding the single best state sequence. This algorithm compensates the global optimal state sequence for test data. However, beam pruning is normally used because full Viterbi decoding is impossible due to its high computational cost and memory requirements. This beam pruning is based on the schema that the local optimal state for each frame corresponds to the global optimal state sequence for the test data. However, this assumption is sometimes broken. Going back to forward decoding, it is based on the local optimal state for each frame. Therefore, combining viterbi decoding and forward decoding that is instead of beam pruning, may result in a better recognition algorithm.