Korea's LILAC codec slashes audio re-encoding waste

Researchers from Seoul National University unveiled LILAC (Idempotent Neural Speech Codec), a breakthrough in audio compression technology that eliminates token corruption during re-encoding cycles. The fully convolutional model processes speech at 24kHz with a 9.375Hz frame rate and ultra-low 0.75kbit/s bitrate while maintaining competitive audio quality (UTMOS 4.14-4.24 on LibriSpeech/LibriTTS-R datasets).

The key innovation lies in LILAC's idempotent architecture, ensuring decoded audio can be re-encoded without altering the original token stream - a first in neural audio codecs. This solves a pervasive problem where traditional systems rewrite 15% of tokens during decode-re-encode cycles, causing quality degradation in multi-stage audio pipelines. The team validated their approach across 12 baseline systems, demonstrating zero token corruption during iterative processing.