TTS (Work in Progress...)

TTS targets a Text2Speech engine lightweight in computation with hight quality speech construction.

Here we have pytorch implementation of Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model as the start point. We plan to improve the model by the recent updated at the field.

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

Requirements

Highly recommended to use miniconda for easier installation.

• python 3.6
• pytorch > 0.2.0
• TODO

Audio Samples

All samples below are generated in test setting and can be generated a model trained on the master branch.

• https://soundcloud.com/user-565970875/tts-ljspeech-val-35000 (iteration 35000)
• https://soundcloud.com/user-565970875/tts-ljspeech-val-13585

Checkpoints

Models	Commit	Audio Sample
iter-62410	99d56f7	https://soundcloud.com/user-565970875/99d56f7-iter62410

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.

You can also enjoy Tensorboard with couple of good training logs, if you point --logdir the experiment folder.

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,
  "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
  "checkpoint": true,  // if save checkpoint per save_step
  "save_step": 200,
  "data_path": "/path/to/KeithIto/LJSpeech-1.0",
  "output_path": "/path/to/my_experiment",
  "log_dir": "/path/to/my/tensorboard/logs/"
}

Testing

Best way to test your pretrained network is to use the Notebook under notebooks folder.

Contribution

Any kind of contribution is highly welcome as we are propelled by the open-source spirit. If you like to add or edit things in code, please also consider to write tests to verify your segment so that we can be sure things are on the track as this repo gets bigger.

TODO

• Make the default Tacotron architecture functional with reasonable fidelity.
• Update the architecture with the latest improvements at the field. (e.g. Monotonic Attention, World Vocoder)
  • Using harmonized teacher forcing proposed by
  • Update the attention module with a monotonic alternative. (e.g GMM attention, Window based attention)
  • References:
    • Efficient Neural Audio Synthesis
    • Attention-Based models for speech recognition
    • Generating Sequences With Recurrent Neural Networks
    • Char2Wav: End-to-End Speech Synthesis
    • VoiceLoop: Voice Fitting and Synthesis via a Phonological Loop
• Simplify the architecture and push the limits of performance vs efficiency.
• Improve vocoder part of the network.
  • Possible Solutions:
    • WORLD vocoder
    • WaveRNN
    • Faster WaveNet
    • Parallel WaveNet

Precursor implementations

• https://github.com/keithito/tacotron (Dataset and Test processing)
• https://github.com/r9y9/tacotron_pytorch (Initial Tacotron architecture)