In this paper we describe a trellis representation of variable length coded data which is capable of being used for bit-level or symbol-level maximum a posteriori (MAP) decoding of variable length codes (VLC). A bit-level soft-input/soft-output module is derived and is applied in an iterative decoding structure consisting of an outer variable length code and an inner convolutional code. Due to their inherent redundancy reversible variable length codes (RVLC) yield good results with this system. We present simulation results in terms of symbol error rate performance when the data is transmitted over a fully interleaved Rayleigh fading channel using BPSK modulation. As measure for the symbol error rate the Levenshtein distance is used which regards the self-synchronizing properties of variable length codes better than a simple symbol-by-symbol comparison.