Scalable and Efficient Neural Speech Coding: A Hybrid Design

We present a scalable and efficient neural waveform coding system for speech compression. formulate the problem as an autoencoding task, where convolutional network (CNN) performs encoding decoding codec (NWC) during its feedforward routine. The proposed NWC also defines quantization entropy trainable module, so artifacts bitrate control are handled optimization process. achieve efficiency by introducing compact model components to NWC, such gated residual networks depthwise separable convolution. Furthermore, models with architecture, cross-module learning (CMRL), cover wide range of bitrates. To this end, we employ concept concatenate multiple modules, each module restore any reconstruction loss that preceding modules have created. CMRL can scale down lower bitrates well, which it employs linear predictive (LPC) first autoencoder. hybrid design integrates LPC redefining LPC's differentiable process, making training end-to-end manner. decoder is either one (0.12 million parameters) in low medium ranges (12 20 kbps) or two NWCs high (32 kbps). Although complexity not yet conventional codecs, significantly reduced from other coders, WaveNet-based vocoder. For wide-band quality, our yields comparable superior performance AMR-WB Opus on TIMIT test utterances at up higher near transparent performance.