Neural Audio Codecs: Variable Time Resolution and Robustness

DAC architecture with SFI convolutional layers controlling neural audio token temporal resolution — Architecture of the variable-token-temporal-resolution neural audio codec (Fig. 2). Nakamura et al., IWAENC 2026

We develop neural audio codecs that adjust token temporal resolution within a single model using sampling-frequency-independent convolutional layers. In a related but distinct line of research, we examine neural audio codec vulnerabilities through white-box adversarial-example generation.

Related Publications

Tomohiko Nakamura, Wataru Nakata, Kanami Imamura, and Yuki Saito, “Neural audio codec with adjustable token temporal resolution using sampling-frequency-independent convolutional layers,” in Proceedings of International Workshop on Acoustic Signal Enhancement, Sept. 2026.