Efficient statistical pruning for maximum likelihood decoding