A Robust Quantized Speech Semantic Communication System Based on ASR Perception Loss and Deep Channel Estimation

Deep-learning-based semantic communication represents a novel paradigm with significant potential to enhance transmission efficiency by focusing on the precise exchange and reconstruction of semantic information. However, existing methods often struggle to mitigate information loss in complex channel environments. To address semantic distortion and severe channel impairments in speech transmission, we propose a robust, quantization-based semantic communication framework. Central to this system is an innovative channel transmission module that employs a soft quantization scheme and a hybrid channel estimator—integrating least-squares (LS) estimation with a U-Net architecture to effectively counteract channel fading. In addition, we introduce an automatic speech recognition (ASR)-aware loss function to simultaneously optimize semantic fidelity and auditory quality. Experimental results demonstrate that our approach significantly outperforms baseline schemes in terms of short-time objective intelligibility (STOI), word error rate (WER), and subjective perceived quality across additive white Gaussian noise (AWGN), Rician, and Rayleigh fading scenarios. Notably, the method achieves an approximately 22% performance improvement under Rayleigh channels at 0-dB signal-to-noise ratio (SNR) while maintaining exceptional transmission efficiency. Furthermore, the proposed framework exhibits favorable characteristics regarding robustness testing and model complexity.