Based on the encoding process, arithmetic codes can be viewed as tree codes and current proposals for decoding arithmetic codes with forbidden symbols belong to sequential decoding algorithms and their variants. However, arithmetic coding can also be modeled as a finite state machine and can be treated as a variable-length trellis code. The number of states used for decoding can be reduced and techniques used for convolutional codes such as the list Viterbi decoding algorithm can be applied on the trellis. The proposed approach provides a rich environment for the design of joint source/channel codes. The particular implementation presented here shows significant performance improvement over previous approaches.