PyTorch implementation of Tacotron: Towards End-to-End Speech Synthesis, and
PyTorch implementation of Natural TTS synthesis by conditioning Wavenet on MEL spectogram predictions.

• Easy switch between Tacotron and Tacotron2
• Detailed model structure configuration with json
  • For Tacotron: tacotron1.json
  • For Tacotron2: tacotron2.json

New configurations can be created by merging features from the following different branches.

• master: Basic Tacotron and Tacotron2 implementation

• dynamic_r: Dynamic reduction factor (r) changing along with training schedule

• gst: Global style token (GST) support

• multispeaker: Multi-speaker support with speaker embeddings

1. Prepare DATASET directory
   • Prepare train.csv.txt and val.csv.txt files
   • Change training_files and validation_files in hparams.py to the above two files respectively
   • Make necessary modifications to files_to_list to retrieve 'mel_file_path' and 'text' in utils/dataset.py
2. Install PyTorch
3. Install python requirements or build docker image
   • Install python requirements: pip install -r requirements.txt

1. python train.py -o outdir -l logdir
2. (OPTIONAL) tensorboard --logdir=logdir

Training using a pre-trained model can lead to faster convergence
By default, the dataset dependent text embedding layers are ignored

1. Download the published Tacotron model
2. python train.py -o outdir -l logdir -c tacotron_statedict.pt --warm_start

1. python train.py -o outdir -l logdir --hparams=distributed_run=True

1. Download published Tacotron model
2. Download published WaveGAN model
3. jupyter notebook --ip=127.0.0.1 --port=31337
4. Load inference.ipynb

Note: When performing Mel-Spectrogram to Audio synthesis, make sure Tacotron and the Mel decoder were trained on the same mel-spectrogram representation.

This implementation uses code from the following repos as described in our code.

• Tacotron2 by NVIDIA
• Tacotron by blue-fish in Real-Time-Voice-Cloning
• Tacotron by r9y9
• Tacotron by keithito