Variable Bitrate Residual Vector Quantization for Audio Coding
Yunkee Chae, Woosung Choi, Yuhta Takida, Junghyun Koo, Yukara Ikemiya, Zhi Zhong, Kin Wai Cheuk, Marco A. Martínez-Ramírez, Kyogu Lee, Wei-Hsiang Liao, Yuki Mitsufuji
TL;DR
This work tackles inefficiency in RVQ-based neural audio codecs caused by a fixed per-frame codebook count, especially for simple frames. It introduces Variable Bitrate RVQ (VRVQ), which uses an importance-map–driven mask to allocate a variable number of codebooks per frame and a smooth gradient estimator for the non-differentiable binarization, trained with a rate-distortion objective. The approach first demonstrates a VRVQ framework on a DAC-based backbone, showing improved rate-distortion metrics and robust scalability to larger codebooks and bitrates, while also enabling multi-bitrate operation without an explicit entropy model. The findings suggest that adaptive per-frame coding and enhanced gradient flow enable more efficient neural audio codecs and hold practical impact for streaming and storage applications.
Abstract
Recent state-of-the-art neural audio compression models have progressively adopted residual vector quantization (RVQ). Despite this success, these models employ a fixed number of codebooks per frame, which can be suboptimal in terms of rate-distortion tradeoff, particularly in scenarios with simple input audio, such as silence. To address this limitation, we propose variable bitrate RVQ (VRVQ) for audio codecs, which allows for more efficient coding by adapting the number of codebooks used per frame. Furthermore, we propose a gradient estimation method for the non-differentiable masking operation that transforms from the importance map to the binary importance mask, improving model training via a straight-through estimator. We demonstrate that the proposed training framework achieves superior results compared to the baseline method and shows further improvement when applied to the current state-of-the-art codec.