Improved MELP speech coding algorithm for neural networks

This paper proposes a KAN-AE autoencoder network, integrating Kolmogorov-Arnold Network (KAN) and Residual Vector Quantizer (RVQ), to improve Mixed Excitation Linear Prediction (MELP) coding for ultra-low bitrate speech compression. The method enhances LSF parameter quantization by extracting semantic-level information, reducing redundancy and improving efficiency. Experimental results show that at a bit rate of 1.96 kbps and LSF quantization below 15 bits, the proposed method achieves a PESQ score of 2.6, outperforming other methods.