BridgeCodec: Mamba Enhanced Neural Audio Codec with Schrödinger Bridge at Low Bitrate

Abstract: Neural audio codecs (NACs) achieve remarkable performance but operate as closed ecosystems requiring strictly co-trained encoder-decoder pairs. This rigid coupling creates a severe interoperability gap in real-world heterogeneous networks, preventing direct communication between devices with codecs deployed independently. To resolve this, we propose BridgeCodec, a versatile framework that fundamentally decouples codec pairs by directly translating highly compressed latent representations between any mismatched, frozen endpoints. We formulate this cross-codec translation as a Schrödinger Bridge optimal transport problem, robustly mapping disparate empirical distributions without predefined noise priors. Additionally, we introduce a Mamba-enhanced U-Net architecture tailored to efficiently capture long-range speech temporal dependencies. To demonstrate its robustness in extreme deployment scenarios, we evaluate BridgeCodec on a highly challenging testbed: mapping a lightweight 1 kbps/8 kHz encoder directly to a 48 kHz decoder. Experimental results show that BridgeCodec achieves superior wideband reconstruction from heavily bottlenecked priors, maintaining high perceptual quality even with a single function evaluation.

Fig. 1