Learned speech codecs have surpassed traditional signal-processing methods in rate-distortion performance, but many of the best neural models rely on sequence-level correlations, which blocks true streaming. The new StreamLSC architecture addresses this by operating frame-by-frame, using components designed for causal, low-latency inference: Boundary-Cached Causal Convolution, Intra-Frame Layer Normalization, and an RNN-mode RWKV module. A streamable entropy quantizer further keeps the context model efficient without violating causality.

In experiments, the streaming model performed comparably to its offline counterpart and achieved state-of-the-art rate-distortion results. Subjective evaluations showed large gains, with average BD-RATE savings of -51.37% on ViSQOL and -53.99% on PESQ, plus improvements of 0.2154 in BD-ViSQOL and 0.5156 in BD-PESQ. These results suggest that frame-wise streaming can be added to high-performance neural codecs without sacrificing quality, which is particularly relevant for real-time and satellite communication scenarios.