TTS (Work in Progress...)

This project is a part of Mozilla Common Voice. TTS targets a Text2Speech engine lightweight in computation with high quality speech synthesis. You might hear a sample here.

Here we have pytorch implementation of Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model. We plan to improve the model by the time with new architectural updates.

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 0.4
• librosa
• tensorboard
• tensorboardX
• matplotlib
• unidecode

Checkpoints and Audio Samples

Checkout here to compare the samples (except the first) below.

Models	Commit	Audio Sample	Details
iter-62410	99d56f7	link	First model with plain Tacotron implementation.
iter-170K	e00bc66	link	More stable and longer trained model.
Best: iter-270K	256ed63	link	Stop-Token prediction is added, to detect end of speech.
Best: [iter-K]	bla	link	Location Sensitive attention

Data

Currently TTS provides data loaders for

• LJ Speech

Training and Finetunning

Split metadata.csv into train and validation subsets respectively metadata_train.csv and metadata_val.csv.

shuf metadata.csv > metadata_shuf.csv
head -n 12000 metadata_shuf.csv > metadata_train.csv
tail -n 11000 metadata_shuf.csv > metadata_val.csv

To train a new model, you need to define a config.json file (simple template below) and call with the command below.

train.py --config_path config.json

To finetune a model, use --restore_path argument.

train.py --config_path config.json --restore_path /path/to/your/model.pth.tar

If you like to use specific set of GPUs, you need set an environment variable. The code uses automatically all the provided GPUs for data parallel training. If you don't specify the GPUs, it uses all GPUs of the system.

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

Each run creates an experiment folder with some meta information, under the folder you set in config.json. In case of any error or intercepted execution, if there is no checkpoint yet under the execution folder, the whole folder is going to be removed.

You can also enjoy Tensorboard, if you point the Tensorboard argument--logdir to the experiment folder.

Example config.json:

{
  "model_name": "my-model", // used in the experiment folder name
  "num_mels": 80,
  "num_freq": 1025,
  "sample_rate": 20000,
  "frame_length_ms": 50,
  "frame_shift_ms": 12.5,
  "preemphasis": 0.97,
  "min_level_db": -100,
  "ref_level_db": 20,
  "embedding_size": 256,
  "text_cleaner": "english_cleaners",

  "epochs": 1000,
  "lr": 0.002,
  "warmup_steps": 4000,
  "batch_size": 32,
  "eval_batch_size":32,
  "r": 5,

  "griffin_lim_iters": 60,
  "power": 1.5,

  "num_loader_workers": 8,

  "checkpoint": true,
  "save_step": 376,
  "data_path": "/my/training/data/path",
  "min_seq_len": 0,
  "output_path": "/my/experiment/folder/path"
}

Testing

Best way to test your pretrained network is to use Notebooks 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 track as this repo gets bigger.

TODO

Checkout issues and Project field.

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
• 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)