Are reconstruction metrics still enough to eval neural audio codecs?

Gauri_the_great · x · 2026-11-19

With neural audio codecs now heavily used in generative speech models, the author questions how long evaluation can keep relying on reconstruction metrics. Most codecs still lean on rate-distortion objectives combining spectral reconstruction, adversarial losses, and quantization objectives — but these say little about the conditional distribution an AR model has to learn, and AR models dominate for speed.

Original post →

More from Research

Research channel →