Test-Time Adaptation via Autoregressive Speech Priors

This method improves speech enhancement under mismatched acoustic conditions by minimizing Kullback-Leibler divergence between enhanced speech and a clean speech prior. The adaptation uses an autoregressive prior trained on clean latent representations from a neural audio codec.