MIREX2018 English Lyrics Alignment

This repository shows our submissions (CW2 and CW3) for subtask 2 of lyrics alignment of MIREX2018. HTK and Tensorflow are used in our submission. More details can be found in our abstract.

Requirements

• FFmpeg
• Git LFS
• HTK 3.4.1
• Python 3.6

Usage

• Clong this repo:

  git clone git@github.com:KKBOX/mirex2018-english-lyrics-alignment.git
  git lfs install
  git lfs pull

• Setup Python virtual environment and install dependences:

  virtualenv -p python3 venv
  source venv/bin/activate
  pip install -r requirements.txt

• Calling format:

  python3 go.py %input_audio %input_txt %output_txt

  • Example:

    python3 go.py example_data/Muse.GuidingLight.mp3 example_data/Muse.GuidingLight.txt output.txt

  • To switch between CW2 and CW3, add --model_dir model_CW2 or --model_dir model_CW3 respectively. The default is model_CW3.

• To use your own model, put the files of configuration (for HCopy), dictionary, list of models, and macro in a directory. Those files should be named mfcc39.edaz.cfg, dic.dic, model_list.model, and macro.final, respectively.

• The running time is about 2.33 min for Muse.GuidingLight.mp3 on a machine with 2.50GHz CPU.

Notes

• Example data came from the Mauch dataset provided in MIREX 2018.