20-clip listening study

Noise in.
Speech restored.

A frozen 10-codebook FishDAC codec, paired with a compact residual Conformer that learns to move noisy decoder latents toward clean ones.

—Noisy mel error
—Enhanced mel error
—Mean reduction
—Checkpoint step

How to listen

Compare each row left to right.

Ground truth is normalized source audio. Noisy is the frozen codec’s reconstruction after white-noise corruption. Enhanced passes those noisy DAC latents through the denoiser before the same frozen decoder. Spectrogram colors share an 80 dB scale within each row.

Loading the listening panel…