A joint source-channel en/decoding scheme based on a new symbol -level joint trellis for the Variable length coding (VLC) is presented. In the encoder part, the serial VLC and a Recursive systematic convolutional (RSC) code are treated as a single component code, which is parallel with another RSC code through a quantizer. According to the joint source-channel en/deco- ding model, we construct a new symbol-level joint trellis with compound states by merging a VLC trellis with a convolutional trellis. Based on this joint trellis, the symbol-level a posteriori probability (APP) decoding algorithm is also derived, which leads to a joint iterative decoding approach with symbol-level soft outputs. The experimental results show that the joint source-channel en/ decoding scheme has obtained better performance than existent joint iterative decoding based on the bit-level super trellis. As a result, the proposed scheme obtains 0.70 dB gains after the 8-th iteration at an SER of 10?4, and it provides substantial error protection for variable-length encoded image data.