MuTr

Transcribe guitar recordings in WAVE format to notes and tablatures (Guitar Pro 5 format).
Pipeline steps:

• Onset detection
• Polyphonic pitch detection
• String and fret detection
• Tempo detection
• Mapping of onset times to discrete notes in measures
• GP5 export

Also offers functionality to compare GP5 files and convert MIDI files to GP5.
The system is implemented in Python 3.5.

Installation

This is currently only tested for Windows 10, but other platforms should work as well.

1. Download and extract this repository.

2. Download and install Anaconda3 64-bit (https://www.continuum.io/downloads).

3. Create and activate a new environment containing all relevant modules using the following commands in an Anaconda prompt:

   conda env create -f $INSTALLDIR/conda_env.yml
   activate music_transcription
   

4. Change the following Keras attributes in the file %USERPROFILE%/.keras/keras.json ($HOME/.keras/keras.json for *nix):

   "image_data_format": "channels_first"
   "backend": "theano"
   

   More info: https://keras.io/backend/

5. (optional) To speed up the CNNs used for onset and pitch detection and if you have a fast GPU, consider running Keras / Theano on the GPU. For even more speed, activate CuDNN and CNMeM. See these two links:
   http://ankivil.com/installing-keras-theano-and-dependencies-on-windows-10/
   http://ankivil.com/making-theano-faster-with-cudnn-and-cnmem-on-windows-10/

   This is recommended if you plan to train new models.

Getting started

Open an Anaconda Prompt and switch the working directory to $INSTALLDIR/pipelines.

Transcribe a recording using polyphonic pitch detection:

python guitar_pipeline.py ..\example_recordings\instrumental_lead.wav

Transcribe a recording using monophonic pitch detection and a custom output path:

python guitar_pipeline.py ..\example_recordings\instrumental_lead.wav -m mono -p instrumental_lead.mono.gp5

Compare two GP5 files:

python compare_gp5.py instrumental_lead.gp5 instrumental_lead.mono.gp5 --outfile instrumental_lead_poly_vs_mono.gp5

The output is another GP5 file with three different tracks, one for common notes and two more describing the notes in the differing regions for each file.

Convert a MIDI file with drums to GP5:

python midi_transcription.py ..\example_recordings\lotrify_laura_drums_4-4.mid --force_drums

To see what other options guitar_pipeline.py has to offer:

python guitar_pipeline.py -h

To see what other options compare_gp5.py has to offer:

python compare_gp5.py -h

To see what other options midi_transcription.py has to offer:

python midi_transcription.py -h