P2PSynCodec neural codec delivers 0.5 kbps speech with 2.0 kbps quality

Most neural speech codecs rely on residual vector quantization (RVQ), where later quantization layers contribute diminishing returns but consume the same bitrate — a fundamental inefficiency. In a new paper accepted to INTERSPEECH 2026, Xiao-Hang Jiang and colleagues introduce P2PSynCodec, which swaps out RVQ for a plain-to-pseudo synergistic vector quantizer (P2PSVQ). The architecture uses a single plain VQ to produce basic quantized tokens, while multiple pseudo VQs generate auxiliary tokens through neural prediction rather than direct quantization. Because the pseudo VQ tokens are predicted from the plain VQ tokens, they incur zero transmitted bitrate, yet they provide enough information for the decoder to reconstruct high-quality speech.

Experimental results show that P2PSynCodec achieves speech reconstruction quality comparable to competing codecs running at 2.0 kbps — but at only 0.5 kbps, a 4x reduction in bitrate. This is a significant leap for ultra-low-bitrate speech coding, with potential applications in bandwidth-constrained environments like satellite communications, hearing aids, and streaming over poor network conditions. The codec maintains robustness across different speakers and languages, suggesting it can be a practical drop-in replacement for existing codecs in resource-limited settings.