Audio sample demo

Attacking UTMOS: Probing the Robustness of a Speech Quality Assessment Model

Wen-Chin Huang and Tomoki Toda

Nagoya University, Japan

Interactive Audio Samples

Click a point on each PESQ vs. UTMOS map to compare the clean reference with the selected sample for three LibriSpeech utterances.

Noise Baseline

Gaussian noise added in waveform, spectrogram, or EnCodec latent space.

PESQ versus UTMOS scatter plot for random noise perturbation baseline

Score-Preserving Attack

Adversarial samples optimized to keep UTMOS high while lowering perceptual quality.

PESQ versus UTMOS scatter plot for score-preserving attacks

Quality-Preserving Attack

Adversarial samples optimized to lower UTMOS while preserving perceptual quality.

PESQ versus UTMOS scatter plot for quality-preserving attacks