TTS (Work in Progress...)

Here we have pytorch implementation of:

• Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model.
• Tacotron2 (TODO): Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions

At the end, it should be easy to add new models and try different architectures.

You can find here a brief note about possible TTS architectures and their comparisons.

Requirements

Highly recommended to use miniconda for easier installation.

• python 3.6
• pytorch > 0.2.0
• TODO

Data

Currently TTS provides data loaders for

• LJ Speech

Training the network

To run your own training, you need to define a config.json file (simple template below) and call with the command.

train.py --config_path config.json

If you like to use specific set of GPUs.

CUDA_VISIBLE_DEVICES="0,1,4" train.py --config_path config.json

Each run creates an experiment folder with the corresponfing date and time, under the folder you set in config.json. And if there is no checkpoint yet under that folder, it is going to be removed when you press Ctrl+C.

Example config.json:

{
  // Data loading parameters
  "num_mels": 80,
  "num_freq": 1024,
  "sample_rate": 20000,
  "frame_length_ms": 50.0,
  "frame_shift_ms": 12.5,
  "preemphasis": 0.97,
  "min_level_db": -100,
  "ref_level_db": 20,
  "hidden_size": 128,
  "embedding_size": 256,
  "text_cleaner": "english_cleaners",

  // Training parameters
  "epochs": 2000,
  "lr": 0.001,
  "lr_patience": 2,  // lr_scheduler.ReduceLROnPlateau().patience
  "lr_decay": 0.5,   // lr_scheduler.ReduceLROnPlateau().factor
  "batch_size": 256,
  "griffinf_lim_iters": 60,
  "power": 1.5,
  "r": 5,            // number of decoder outputs for Tacotron

  // Number of data loader processes
  "num_loader_workers": 8,

  // Experiment logging parameters
  "save_step": 200,
  "data_path": "/path/to/KeithIto/LJSpeech-1.0",
  "output_path": "/path/to/my_experiment",
  "log_dir": "/path/to/my/tensorboard/logs/"
}