CERBERUS Network for Music Separation & Transcription

---

This is a rough implementation of Simultaneous Separation and Transcription of Mixtures with Multiple Polyphonic and Percussive Instruments (Ethan Manilow et al., ICASSP2020)

Note

This implementation did not achieve as much performance as reported in the paper.

Demo (Source Separation)

Quantitative Evaluation (Transcription)

	Precision	Recall	Accuracy
Piano	0.585	0.566	0.460
Bass	0.797	0.817	0.747
Drums	0.230	0.417	0.133

• Note : There's no benchmark dataset. These results are measured on data I randomly created using test set of Slakh2100 dataset. So It is not appropriate to quantitatively compare these results with those reported in the paper.

Pretrained Network & config

• Weight(ckpt)
• Config(yaml)

Inference

python inference.py hparams.yaml weight.ckpt input.wav output_dir/

• Expected Results:

  

Training with Slakh2100 Dataset

1. Get Slakh2100 dataset (See: Slakh2100 Project)
2. Downsample audio to 16k
3. Modify configs/config.yaml

   data_dir: "/path/to/slakh2100_flac_16k/"
   
   # see: validation_epoch_end() in network/cerberus_wrapper.py
   sample_audio:
   path: "/path/to/sample/audio/sample_rate_16k.wav"
   offset: 1264000 
   num_frames: 160000

4. Run training

   python train.py

Contact

• Jongho Choi ([email protected])
• Jiwon Kim
• Ahyeon Choi