aiqbench analysis
This paper introduces LILAC, a novel neural speech codec designed to solve the idempotency problem where standard systems alter audio tokens upon re-encoding. By constructing a fully convolutional architecture that preserves token streams through decode-re-encode cycles, it enables more robust pipeline integration for low-bitrate applications like 9.375 Hz compression at 0.75 kbit/s. The approach maintains competitive quality metrics comparable to current state-of-the-art models while addressing a critical structural limitation in existing neural audio codecs.
LILAC: An Idempotent Neural Speech Codec
arXiv:2608.05727v1 Announce Type: cross Abstract: Neural Audio Codecs are widely adopted in speech generation and editing. However, existing neural audio codecs are not idempotent: across the paper's twelve baseline systems, every configuration tested rewrites, on average, at least 15% of its tokens in a single decode-re-encode pass. This poses a problem for utilizing Neural Audio Codecs as token interfaces in pipelines where re-encoding decoded outputs can occur. We present LILAC, a fully convolutional 24 kHz speech codec at 9.375 Hz and 0.75 kbit/s that is codec idempotent by construction; re-encoding the decoded audio of any valid token stream returns the identical stream. LILAC achieves idempotency while maintaining competitive quality, reaching UTMOS 4.14 and 4.24 on LibriSpeech and LibriTTS-R test sets, comparable to SOTA sub-1 kbit/s Neural Audio Codecs.
Abstract:Neural Audio Codecs are widely adopted in speech generation and editing. However, existing neural audio codecs are not idempotent: across the paper's twelve baseline systems, every configuration tested rewrites, on average, at least 15% of its tokens in a single decode-re-encode pass. This poses a problem for utilizing Neural Audio Codecs as token interfaces in pipelines where re-encoding decoded outputs can occur. We present LILAC, a fully convolutional 24 kHz speech codec at 9.375 Hz and 0.75 kbit/s that is codec idempotent by construction; re-encoding the decoded audio of any valid token stream returns the identical stream. LILAC achieves idempotency while maintaining competitive quality, reaching UTMOS 4.14 and 4.24 on LibriSpeech and LibriTTS-R test sets, comparable to SOTA sub-1 kbit/s Neural Audio Codecs.
Sound (cs.SD); Machine Learning (cs.LG)
arXiv:2608.05727 [cs.SD]
(or arXiv:2608.05727v1 [cs.SD] for this version)
https://doi.org/10.48550/arXiv.2608.05727 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 219 words
Related coverage
How Robust Are Neural Audio Codecs for African Speech? A Multi-Task Benchmark and the Limits of Perceptual Quality
arXiv · Oct 2, 2026
MyoCodec: A Streaming Neural Codec for Electromyography
arXiv · Sep 30, 2026
CodecFlow: Efficient Bandwidth Extension via Conditional Flow Matching in Neural Codec Latent Space
arXiv · Oct 15, 2026
Assessing speech quality metrics for evaluation of neural audio codecs under clean speech conditions
arXiv · Aug 5, 2026