ZeroSC-SR: Importance-Aware Zero-Shot Semantic Speech Communication Over Fading Channels

This paper introduces an innovative Zero-Shot Semantic Communication System with Speech Reconstruction (ZeroSC-SR), designed to enhance speech transmission and reconstruction in frequency-selective fading channels while remaining compatible with standard digital coding and modulation. By efficiently transmitting phoneme sequences and compressed prompt acoustic features as discrete indices, ZeroSC-SR reduces transmitted data size and achieves a high compression ratio, resulting in lower Character Error Rate (CER) and higher speaker similarity at low transmit power under noisy channels. We quantitatively evaluate the semantic importance of the transmitted phoneme sequences and acoustic features through a task-oriented, mutual-information-based approach. These insights directly guide the development of our Iterative Importance-Aware Allocation and Power Scheduling (IIAPS) strategy, which enhances transmission quality in frequency selective fading channels by allocating subcarriers and power based on semantic importance. Experimental results show that, under fixed power with SNR sweeps, IIAPS outperforms baseline allocation strategies on Word Error Rate (WER) and speaker similarity. Compared to conventional systems, ZeroSC-SR de livers lower CER and higher speaker similarity under varying transmit power, with the greatest improvements occurring at lower power levels.