Semantic-Preserved Communication System for Highly Efficient Speech Transmission

Deep learning (DL) based semantic communication methods have been explored for the efficient transmission of images, text, and speech in recent years. In contrast to traditional wireless that focus on abstract symbols, approaches attempt achieve better efficiency by only sending semantic-related information source data. this paper, we consider semantic-oriented which transmits semantic-relevant over channel recognition task, a compact additional set semantic-irrelevant reconstruction task. We propose novel end-to-end DL-based transceiver extracts encodes from input spectrums at transmitter outputs corresponding transcriptions decoded receiver. particular, employ soft alignment module redundancy removal extract text-related features while dropping semantically redundant content, greatly reducing amount compared existing methods. also correction further correct predicted transcription with knowledge leveraging pretrained language model. For transmission, include CTC small number but speech-related information, such as duration, pitch, power speaker identification original signals introduce two-stage training scheme speeds up proposed DL The simulation results confirm our method outperforms current terms accuracy text quality recovered significantly improves efficiency. More specifically, sends 16% transmitted symbols required achieving about 10% reduction WER transmission. it an even more remarkable improvement 0.2% preserving comparable reconstructed signals.