• Implementation of "Neural Speech Synthesis with Transformer Network"
• This is implemented for FastSpeech

1. Download and extract the LJ Speech dataset
2. Make preprocessed folder in LJSpeech directory and make char_seq & phone_seq & melspectrogram folder in it
3. Set data_path in hparams.py as the LJSpeech folder
4. Using prepare_data.ipynb, prepare melspectrogram and text (converted into indices) tensors.
5. python train.py

• Stop prediction loss (train / val)
  
• Guided attention loss (train / val)
  
• L1 loss (train / val)
  

• Encoder Alignments

You can hear the audio samples here

1. Unlike the original paper, I didn't use the encoder-prenet following espnet
2. I apply additional "guided attention loss" to the two heads of the last two layers
3. Batch size is important, so I use gradient accumulation
4. You can also use DataParallel. Change the n_gpus, batch_size, accumulation appropriately.

1. Dynamic batch

1. For fastspeech, generated melspectrograms and attention matrix should be saved for later.
   1-1. Set teacher_path in hparams.py and make alignments and targets directories there.
   1-2. Using prepare_fastspeech.ipynb, prepare alignmetns and targets.

2. To draw attention plots for every each head, I change return values of the "torch.nn.functional.multi_head_attention_forward()"

#before
return attn_output, attn_output_weights.sum(dim=1) / num_heads  

#after  
return attn_output, attn_output_weights

3. Among num_layers*num_heads attention matrices, the one with the highest focus rate is saved.

1.NVIDIA/tacotron2: https://github.com/NVIDIA/tacotron2
2.espnet/espnet: https://github.com/espnet/espnet
3.soobinseo/Transformer-TTS: https://github.com/soobinseo/Transformer-TTS