nbb5858 / wavenet_implementation Goto Github PK

This is a TensorFlow implementation of the WaveNet model first introduced in WaveNet: A Generative Model for Raw Audio. The architecture for the model is based on appendix A of Deep Voice: Real-time Neural Text-to-Speech. The scripts within are tested on TensorFlow 2.13.0 and Python 3.9.6.

The weights for a demo model are saved at Saved_Models >> Demo_Model. This model was trained on ~60 minutes of Bach's Well-Tempered Klavier. A demo result for this model is located at Generated_Samples >> demo_sample.wav.

To train the network,

1. Place the training set into the Training_Data folder and the validation set in to the Validation_Data folder. Training & validation data should be .wav files sampled at 8kHz.

2. In params.json, define the network parameters. The defaults are those used in the demo model.

3. Navigate to train.py. Define the number of epochs, n_epochs. If this is the first run at training the network, set load_from_weights = False. The training script will save the weights for the most recent model at Saved_Models >> Last_Model. It will save the weights from the best model, i.e., with the lowest validation loss, at Saved_Models >> Best_Model. Training metrics are saved in the Metrics folder. Upon subsequent training runs, set load_from_weights = True. This will load last set of model weights, so training can be resumed where it was left off.

To generate audio, navigate to the generate.py script. Choose the number of samples to generate num_generate. Choose which validation file, val_file, to use as a seed. Choose the start and end range for the seed, start and end.