ETH Zurich researchers improve audio fidelity using geometric iterative retrieval

This development addresses a critical bottleneck in neural audio codecs where resynthesis quality often limits the fidelity of generated audio. By treating the RVQ hierarchy as a natural iterative axis rather than using external noise schedules, the method provides a more semantically grounded approach to audio reconstruction than current diffusion models. For the streaming industry, this suggests a path toward lower-bitrate audio that maintains high perceptual quality for music and speech applications. The findings also highlight a growing divergence between spectral metrics like LSD and human preference. Watch for whether this geometric approach is integrated into future versions of the Descript Audio Codec or similar B2B audio processing stacks.

The Descript Audio Codec (DAC) has become a reference implementation for neural audio compression research since its open-source release. In 2023, Descript researchers published DAC as a 44.1 kHz neural codec achieving high-fidelity reconstruction at 8 kbps, positioning it as a competitive alternative to EnCodec and SoundStream for both speech and music. The codec's architecture, which uses a multi-scale STFT discriminator and residual vector quantization with up to 32 codebook layers, has been adopted as a baseline in dozens of subsequent papers on audio generation and compression. ETH Zurich's geometric iterative retrieval work builds directly on this RVQ structure, treating the codebook hierarchy as an iterative refinement axis rather than relying on external diffusion schedules.

The broader neural audio codec market is being shaped by both academic competition and commercial deployment. Meta's EnCodec, released in late 2022, established the RVQ-based paradigm that DAC and subsequent codecs refined. In 2025, Google DeepMind introduced SoundStream variants optimized for real-time streaming at bitrates as low as 3 kbps, targeting conversational AI and telephony use cases where latency constraints are tight. Meanwhile, the streaming industry's interest in neural codecs is driven by bandwidth economics: at scale, even a 20% reduction in audio bitrate translates to significant CDN cost savings for platforms serving billions of hours annually. Descript itself has integrated DAC into its commercial audio editing and transcription platform, demonstrating that neural codecs can move from research artifacts into production workflows.

Technical benchmarks from the ETH Zurich paper show that geometric iterative retrieval achieves lower log-spectral distance scores than both discrete token prediction and one-step regression baselines across speech and music datasets. This aligns with a broader trend in codec research where iterative refinement methods are outperforming single-pass approaches. The CLIP-style contrastive loss used to align intermediate representations with target audio features draws on techniques popularized in vision-language models, representing a cross-modal transfer of training objectives into the audio domain. The finding that spectral metrics like LSD diverge from human preference judgments echoes similar observations in video codec evaluation, where VMAF and perceptual quality scores often disagree with PSNR. For streaming applications, this suggests that future neural audio codecs may need to optimize for perceptual metrics rather than purely spectral ones, a challenge that mirrors the video industry's ongoing debate over , a challenge that mirrors the video industry's ongoing debate over objective quality measurement.