TIMUSE

Text-Initialized Masked Semantic Token Modeling for Universal Speech Enhancement

Some examples from the multilingual URGENT 2026 non-blind test set.

For a fair comparison, all columns use lossless mono 24 kHz FLAC files. Headphones are recommended. Mel spectrograms can be shown with the switch below; the white line follows playback.
For this subjective listening demo, 16 kHz baselines such as LLaSE-G1 and UniSE are omitted because differences in sampling rate can noticeably affect perceived audio quality. This does not affect the objective evaluation reported in the paper, where all metrics are computed on 16 kHz resampled audio.
Note that the audio in the “Clean reference” column is taken directly from URGENT 2026. Although used as the source signal for degradation simulation, it is not necessarily perceptually clean.

Comparison using the officially released checkpoints for all baselines. Since the systems differ in training data, degradation simulation, and training recipes, this is not a strictly controlled method-level comparison, but rather an overall indication of their relative performance.

Comparison under the same training dataset and recipe, isolating the effect of model design from differences in training data and degradation simulation.