ParakeetRebeccaRosario/examples/fastspeech/configs/ljspeech.yaml

34 lines
1.4 KiB
YAML

audio:
num_mels: 80 #the number of mel bands when calculating mel spectrograms.
n_fft: 2048 #the number of fft components.
sr: 22050 #the sampling rate of audio data file.
hop_length: 256 #the number of samples to advance between frames.
win_length: 1024 #the length (width) of the window function.
power: 1.2 #the power to raise before griffin-lim.
network:
encoder_n_layer: 6 #the number of FFT Block in encoder.
encoder_head: 2 #the attention head number in encoder.
encoder_conv1d_filter_size: 1536 #the filter size of conv1d in encoder.
max_seq_len: 2048 #the max length of sequence.
decoder_n_layer: 6 #the number of FFT Block in decoder.
decoder_head: 2 #the attention head number in decoder.
decoder_conv1d_filter_size: 1536 #the filter size of conv1d in decoder.
hidden_size: 384 #the hidden size in model of fastspeech.
duration_predictor_output_size: 256 #the output size of duration predictior.
duration_predictor_filter_size: 3 #the filter size of conv1d in duration prediction.
fft_conv1d_filter: 3 #the filter size of conv1d in fft.
fft_conv1d_padding: 1 #the padding size of conv1d in fft.
dropout: 0.1 #the dropout in network.
outputs_per_step: 1
train:
batch_size: 32
learning_rate: 0.001
warm_up_step: 4000 #the warm up step of learning rate.
grad_clip_thresh: 0.1 #the threshold of grad clip.
checkpoint_interval: 1000
max_epochs: 10000