tacotron_hparams.py

# Attention parameters
attention_rnn_dim=1024,  # sets the number of units in the RNN
attention_dim=128,  # sets the number of units in the attention mechanism
#  These two values are relatively large and may require a significant amount of GPU memory during training and inference.

# Location Layer parameters
attention_location_n_filters=32,  # sets the number of filters in the CNN
attention_location_kernel_size=31,  # sets the size of the filters
# This means that the CNN has 32 filters and each filter has a kernel size of 31.

# via https://github.com/tikhonova/what_would_alan_watts_say/blob/master/speech_synthesis/tacotron2_waveglow_model_files/tacotron_hparams.py