A machine-learning model to automatically transcribe audio recordings of solo acoustic guitars to readable guitar tablature built using TensorFlow. Undertaken for my Senior-Honours dissertation at the University of St Andrews, this model builds on academic work in the nascent field of using machine learning for Music Information Retrieval (MIR), particularly the TabCNN model and utilises a convolutional neural network to learn features from spectograms of guitar audio recordings to automatically produce usable guitar tablature
Supervised by Dr Oggie Arandelovic who awarded me a grade of a first and called this project "very good work with some genuine contributions to the state of the art".
To run the project and train the model yourself a few step must be taken.
- The dataset can be downloaded here. Make sure to unzip this folder and place it in the root directory of the project.
- After that, you must activate a virtual environment to install the dependencies. This can be done with
python3 -m venv <env> - Then you can activate the virtual environment by running
source <env>/bin/activate - Next, please install the dependencies with
pip install -r requirements.txt - After this, you have to generate the CQT representations for the audio files using
python3 ParallelGenerateCQTs.py - And finally you can train the model with
python3 PolyTab.py - Once the model has trained, you can run
python3 PolyTabPredictor.py --weights "path/to/weights.h5" --audio "path/to/audio/file.wav"with the trained weights and the audio you want to predict for. The saved predictions can be found in the /predictions folder. - You can then run
python3 PolyTabPredictor.py --weights "saved/c 2024-03-21 171741/5/weights.h5" --audio "path/to/audio/file.wav"to predict using the model which was trained with the learnable weighted loss and AdamW optimiser.
The accompanying paper for this project can be found here