facebookresearch / demucs

Code for the paper Hybrid Spectrogram and Waveform Source Separation
MIT License
8.37k stars 1.07k forks source link

Trumpets in music cannot be separated correctly! #598

Open piwawa opened 6 months ago

piwawa commented 6 months ago

I used the following python script:

from audio_separator.separator import Separator

# Initialize the Separator with other configuration properties below
separator = Separator(output_dir='output')

# Load a model
separator.load_model(model_filename='htdemucs_ft.yaml')

output_file_paths_3 = separator.separate('../spleeter-web/media/uploads/72239f71-f702-4951-9bec-169a1df905af/LeLe Farley feat. DeDe Harlan - Earths Afterparty.webm')

Model:

        "Demucs v4: htdemucs_ft": {
            "04573f0d-f3cf25b2.th": "https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/04573f0d-f3cf25b2.th",
            "92cfc3b6-ef3bcb9c.th": "https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/92cfc3b6-ef3bcb9c.th",
            "d12395a8-e57c48e6.th": "https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/d12395a8-e57c48e6.th",
            "f7e0c4bc-ba3fe64a.th": "https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/f7e0c4bc-ba3fe64a.th",
            "htdemucs_ft.yaml": "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/htdemucs_ft.yaml"
        }

Origin audio:

LeLe Farley feat. DeDe Harlan - Earths Afterparty.webm

Vovals audio:

LeLe Farley feat. DeDe Harlan - Earths Afterparty_(Vocals)_htdemucs_ft.webm

The horn sound at 26 seconds was incorrectly preserved.