Tacotron2 for Mandarin

1. python3 prepare_data.py
2. python3 train.py

Structure

Embedding(256) Encoder: 3 * (conv1d(size=5, out_dim=256, stride=1, dilation=1 + Bn + relu + dropout) Bilstm(128)

Decoder: Prenet: Linear(256) + relu + dropout, Linear(256) + relu + dropout Attention_RNN() + AttentionLayer() + decoder_RNN() PostNet: Conv1d + BN + tanh: Mel_dim -> postnet_embedding_dim 5 * (Conv1d + BN + tanh): postnet_embedding_dim -> postnet_embedding_dim Conv1d + BN: postnet_embedding_dim -> n_mel_channels

Loss: MSE(mel_out, mel_target) + MSE(mel_out_postnet, mel_targets) BCE(gate_output, gate_target) 1. gate_target为预测结束的标志。