A GAN Based Codec with Vocal Tract Features for Low Rate Speech Coding

In this paper, we propose a GAN-based codec for wideband speech coding at low bitrates, which contains a conventional encoder and a neural vocoder with lower dimensional input. The method contains the following two innovations: Firstly, we add vocal tract features as vocoder input, fusing them with energy features, so that the model improves the generated speech quality under limited input dimensions. Secondly, we propose a sub-band time-frequency discriminator. The discriminator divides the input frequency band into several sub-bands according to the auditory sense, independently learning and discriminating the features in each sub-band. The experimental results show that the proposed method reaches the state-of-the-art for low bit rate speech coding, and the computational complexity of proposed method is about 2 GMACs, demonstrating the superior performance of the proposed method.