Audio & Speech

VoCodec cuts speech bitrate by 27% with voicing-driven quantization

New neural speech codec hits 1.1 kbps while outperforming baselines on LibriTTS.

Deep Dive

Traditional neural speech codecs apply uniform quantization across all frames, wasting bits on less perceptually important unvoiced segments. VoCodec, developed by Xiao-Hang Jiang and colleagues, tackles this by embedding a voicing detector into a fully causal encoder-quantizer-decoder framework. Voiced frames receive high-fidelity residual scalar-vector quantization, while unvoiced frames use simpler scalar quantization. This dynamic allocation preserves quality where it matters most, dramatically lowering overall bitrate without sacrificing intelligibility.

On the LibriTTS dataset at 16 kHz sampling, VoCodec outperforms existing neural speech codecs even at an ultra-low bitrate of 1.1 kbps. Further experiments confirm a 27% bitrate reduction compared to uniform quantization strategies. Accepted to INTERSPEECH 2026, VoCodec offers a practical path for real-time, low-bitrate speech applications like VoIP, satellite communication, and hearing aids, where bandwidth is scarce but voice clarity is critical.

Key Points
  • VoCodec uses a voicing detector to adaptively allocate higher bitrate to voiced frames and lower to unvoiced frames based on perceptual sensitivity.
  • Achieves 1.1 kbps bitrate while outperforming baseline neural speech codecs on LibriTTS at 16 kHz sampling rate.
  • Reduces bitrate by approximately 27% compared to uniform quantization strategies, with fully causal streamable architecture.

Why It Matters

Enables high-quality speech transmission at ultra-low bitrates, saving bandwidth for streaming, VoIP, and mobile communications.

📬 Get the top 10 AI stories daily