This will probably be the final release, except for bug fixes.
Thanks to all who have expressed interest in vocal-remover!!!

Note

• Add evaluation script
• Refactor

Note

• Phase estimation with complex mask
• Update the model architecture
• Update the pre-trained model

Evaluation result on DSD100

SDR in dB

Note

• Phase estimation with complex mask
• Update the model architecture
• Update the pre-trained model

Evaluation result on DSD100

SDR in dB

Note

• baseline.pth now can be loaded from anywhere (https://github.com/tsurumeso/vocal-remover/pull/138)
• Update requirements.txt

[!WARNING]
This is a beta release.

Note

• Implement bounded mask
• Update the pre-trained model

Note

• Fix a bug that causes a large disk load on online data loading

[!WARNING]
This is a beta release.

Note

• Phase estimation with complex mask
• Update model architecture
• Update the pre-trained model
• Fix a bug that causes a large disk load on online data loading

Evaluation result on DSD100

Instruments

Vocals

Note

• Add resampy to requirements.txt (#146)

Note

• Update requirements.txt (#136)

Note

• Update link to the paper (#113)
• Add support Apple chip device (#121)
• Updated librosa function calls that do not comply to PEP3102 (#125)

Note

• Parameterize model size
• Add --output_dir option (#100)

Note

• Fix API issue with the latest version of librosa (#99)

Note

• Speed up training and inference
• Reduce memory usage during training
• Update model architecture
• Update the pre-trained model

Note

• Update model architecture
• Update the pre-trained model

Note

• Update training parameters
• Update a post-processing parameter
• Update Usage section of the README

Note

• Hotfix to the training script
• Split --max_reduction_rate option into --reduction_rate and --reduction_level

Note

• Hotfix to the training script

Note

• Implement train/val splitting based on subdirectories
• Update training parameters

Note

• Fix normalization of complex spectrograms
• Make Test Time Augmentation (TTA) optional
• Support for training on CPU

If you want to improve the vocal removal quality with TTA, use the --tta option.

# CPU
python inference.py --input path/to/an/audio/file --tta

# GPU
python inference.py --input path/to/an/audio/file --tta --gpu 0

Note

• Update the pre-trained model
• Remove offline augmentation
• Make voсals removal more aggressive
• Fix #35