Efficient And Scalable Neural Residual Waveform Coding With Collaborative Quantization
This work addresses the need for efficient and scalable neural speech codecs for applications on various devices, representing an incremental improvement by integrating neural networks with traditional signal processing methods.
The paper tackles the problem of scalability and efficiency in neural speech codecs by proposing a collaborative quantization (CQ) scheme that jointly learns codebooks for LPC coefficients and residuals, achieving higher quality than its predecessor at 9 kbps with lower complexity and outperforming AMR-WB and Opus at up to 24 kbps with models under 1 million parameters.
Scalability and efficiency are desired in neural speech codecs, which supports a wide range of bitrates for applications on various devices. We propose a collaborative quantization (CQ) scheme to jointly learn the codebook of LPC coefficients and the corresponding residuals. CQ does not simply shoehorn LPC to a neural network, but bridges the computational capacity of advanced neural network models and traditional, yet efficient and domain-specific digital signal processing methods in an integrated manner. We demonstrate that CQ achieves much higher quality than its predecessor at 9 kbps with even lower model complexity. We also show that CQ can scale up to 24 kbps where it outperforms AMR-WB and Opus. As a neural waveform codec, CQ models are with less than 1 million parameters, significantly less than many other generative models.