Vocal Melody Extraction

This repository includes the source code of the melody extraction algorithm from:

Wei-Tsung Lu and Li Su, “Vocal melody extraction with semantic segmentation and audio-symbolic domain transfer learning,” International Society of Music Information Retrieval Conference (ISMIR), September 2018.

Dependencies

This repository requires following packages:

python 3.6
numpy
tensorflow
keras
mido

Usage

usage: VocalMelodyExtraction.py [-h][-p phase]
                                [-t model_type][-d data_type][-da dataset_path][-la label_path]
                                [-ms model_path_symbolic][-w window_width][-b batch_size_train][-e epoch]
                                [-n steps][-o output_model_name]
                                [-m model_path] [-i input_file][-bb batch_size_train]
  required arguments:
  -da dataset_path              path to data set 
  -la label_path                path to dataset label
  -ms model_path_symbolic       path to symbolic model 
  
  optional arguments:
  -h                
  -p  phase                     phase: training or testing (default: "testing) 
  -t  model_type                model type: seg or pnn (default: "seg")
  -d  data_type                 data type: audio or symbolic (default: "audio") 
  -w  window_width              width of the input feature (default: 128)
  -b  batch_size_train          batch size during training (default: 12)
  -e  epoch                     number of epoch (default: 5)
  -n  steps                     number of step per epoch (default: 6000)
  -o  output_model_name         name of the output model (default: "out")
  -m  model_path                path to existing model (default: "Seg")
  -i  input_file                path to input file (default: "train01.wav")
  -bb batch_size_train          batch size during testing (default: 10)

Pretrained Models

Click here to download the pretrained models.

Todos

Add codes for symbolic model training
Data set handling

License