Cognitive radio (CR) has become one promising solution to the shortage of spectrum resources and transform domain communication system (TDCS) has been proposed as one candidate transmission technique of CR, especially as a transmission candidate of signaling in the CR context. While in most modern communication systems soft-demodulation is concatenated with soft-decision decoding to achieve high performances, current TDCS cannot output soft information. In this work, the soft-demodulation algorithm of TDCS signals in fading channels and its efficient IFFT-based implementation are proposed. This implementation only introduces negligible complexity increase in the receiver; however, it brings performance improvement and makes it possible to efficiently and iteratively soft-decode coded data. Simulation results of both perfect and practical channel estimation in the IEEE 802.22 Profile C channel show that the proposed method outperforms the hard-decision demodulation method, especially in the high-SNR-region.