MISRNet: Lightweight Neural Vocoder Using Multi-Input Single Shared Residual Blocks

Abstract

Neural vocoders have recently become popular in text-to-speech synthesis and voice conversion, increasing the demand for efficient neural vocoders. One successful approach is HiFi-GAN, which archives high-fidelity audio synthesis using a relatively small model. This characteristic is obtained using a generator incorporating multi-receptive field fusion (MRF) with multiple branches of residual blocks, allowing the expansion of the description capacity with few-channel convolutions. How-ever, MRF requires the model size to increase with the number of branches. Alternatively, we propose a network called MISRNet , which incorporates a novel module called multi-input single shared residual block (MISR) . MISR enlarges the description capacity by enriching the input variation using lightweight convolutions with a kernel size of 1 and, alternatively, reduces the variation of residual blocks from multiple to single. Because the model size of the input convolutions is significantly smaller than that of the residual blocks, MISR reduces the model size compared with that of MRF. Furthermore, we introduce an implementation technique for MISR, where we accelerate the processing speed by adopting tensor reshaping. We experimentally applied our ideas to lightweight variants of HiFi-GAN and iSTFTNet, making the models more lightweight with comparable speech quality and without compromising speed. 1