A learned filter
Humpback-CLEAN uses a small U-Net to predict a mask over a recording's time-frequency representation. The mask can attenuate or amplify existing energy, and the waveform is reconstructed with the original phase.
Convert audio into a spectrogram: frequency against time.
Estimate how strongly to retain each time-frequency cell.
Rebuild the waveform, then listen against the unchanged original.
The serving model processes audio at 16 kHz. It does not recover information above 8 kHz. Its bounded mask is not a guarantee against artifacts or distortion, and it should not replace the original recording in scientific analysis.
Hear the difference
Choose a version, press Play, and switch while listening. Switching keeps your playback position. Within each example, all three spectrograms use the same color scale and show 0-4 kHz. Audio is delivered as MP3; the Listen workbench retains access to the original recording.
Loading example recordings...
We could not load the examples. The Listen workbench remains available.
How it was trained
The historical training approach mixed real humpback recordings with ocean noise and trained the model to recover the song target. Some targets were preprocessed rather than pristine clean recordings. Here, “synthetic” refers to the mixture, not generated whale song.
The source recordings span multiple locations, but segments from one recording are not independent animals or encounters. Dataset lineage and encounter-separated evaluation matter more than a large segment count. The Dataset page distinguishes the available curated collections from the full training corpus.
A comparison, not a ranking
Earth Species Project's Biodenoise is a general animal-vocalization denoiser and is also available in OceanSong. Each example here includes its full-strength output, with no preceding noise-reduction stage. Humpback-CLEAN is applied independently to the same original.
These examples do not establish that one system is better overall. That requires held-out recordings, appropriate reference annotations, distortion checks, and a fixed downstream task. Preserve the original, listen for lost calls as well as removed noise, and do not interpret a model score as a calibrated identification probability.
Reference: Miron et al., Biodenoising: Animal Vocalization Denoising without Access to Clean Data.
Try other recordings
The Listen workbench offers Humpback-CLEAN, ESP Biodenoise, noise reduction and original playback. Public hydrophones can go offline, and archived audio is not a live stream.