Cross-Scale Vector Quantization for Scalable Neural Speech Coding
This addresses the need for scalable audio coding in real-time communications, reducing memory footprint and transcoding requirements, though it is incremental as it builds on existing neural coding networks.
The paper tackles the problem of bitrate scalability in neural audio codecs by introducing a cross-scale scalable vector quantization scheme (CSVQ) that encodes multi-scale features progressively, allowing reconstruction from partial bitstreams and quality improvement with more bits. Results show CSVQ at 3 kbps outperforms Opus at 9 kbps and Lyra at 3 kbps, providing a graceful quality boost with bitrate increase.
Bitrate scalability is a desirable feature for audio coding in real-time communications. Existing neural audio codecs usually enforce a specific bitrate during training, so different models need to be trained for each target bitrate, which increases the memory footprint at the sender and the receiver side and transcoding is often needed to support multiple receivers. In this paper, we introduce a cross-scale scalable vector quantization scheme (CSVQ), in which multi-scale features are encoded progressively with stepwise feature fusion and refinement. In this way, a coarse-level signal is reconstructed if only a portion of the bitstream is received, and progressively improves the quality as more bits are available. The proposed CSVQ scheme can be flexibly applied to any neural audio coding network with a mirrored auto-encoder structure to achieve bitrate scalability. Subjective results show that the proposed scheme outperforms the classical residual VQ (RVQ) with scalability. Moreover, the proposed CSVQ at 3 kbps outperforms Opus at 9 kbps and Lyra at 3kbps and it could provide a graceful quality boost with bitrate increase.