Implementation of a non-autoregressive Transformer based neural network for Text-to-Speech (TTS).
This repo is based, among others, on the following papers:

• Neural Speech Synthesis with Transformer Network
• FastSpeech: Fast, Robust and Controllable Text to Speech
• FastSpeech 2: Fast and High-Quality End-to-End Text to Speech
• FastPitch: Parallel Text-to-speech with Pitch Prediction

Our pre-trained LJSpeech model is compatible with the pre-trained vocoders:

• MelGAN
• HiFiGAN

(older versions are available also for WaveRNN)

For quick inference with these vocoders, checkout the Vocoding branch

Being non-autoregressive, this Transformer model is:

• Robust: No repeats and failed attention modes for challenging sentences.
• Fast: With no autoregression, predictions take a fraction of the time.
• Controllable: It is possible to control the speed and pitch of the generated utterance.

Can be found here.

These samples' spectrograms are converted using the pre-trained MelGAN vocoder.

Try it out on Colab:

Hifigan	MelGAN

• 06/20: Added normalisation and pre-trained models compatible with the faster MelGAN vocoder.
• 11/20: Added pitch prediction. Autoregressive model is now specialized as an Aligner and Forward is now the only TTS model. Changed models architectures. Discontinued WaveRNN support. Improved duration extraction with Dijkstra algorithm.
• 01/21: Added model and prediction code for pre-trained MelGAN and HiFiGAN vocoders.
• 03/21: Vocoding branch.

• Installation
• API
• Dataset
• Training
  • Aligner
  • TTS
• Prediction
• Model Weights

Make sure you have:

• Python >= 3.6

Install espeak as phonemizer backend (for macOS use brew):

sudo apt-get install espeak

Then install the rest with pip:

pip install -r requirements.txt

Read the individual scripts for more command line arguments.

If you intend to use the pre-trained vocoders, also install the extra requirements under vocoding

pip install -r vocoding/extra_requirements.txt

Use our pre-trained model (Griffin-Lim, HiFiGAN or MelGAN) from command line with

python predict_tts_vocoder.py -t "Please, say something." --hifigan

Or in a python script

from data.audio import Audio
from model.factory import tts_ljspeech

model, config = tts_ljspeech()
audio = Audio(config)
out = model.predict('Please, say something.')

# Convert spectrogram to wav (with griffin lim)
wav = audio.reconstruct_waveform(out['mel'].numpy().T)

You can directly use LJSpeech to create the training dataset.

• If training on LJSpeech, or if unsure, simply use config/session_paths.yaml to create MelGAN compatible models
  • swap data_config.yaml for data_config_wavernn.yaml to create models compatible with WaveRNN
• EDIT PATHS: in config/session_paths.yaml edit the paths to point at your dataset and log folders

Prepare a folder containing your metadata and wav files, for instance

|- dataset_folder/
|   |- metadata.csv
|   |- wavs/
|       |- file1.wav
|       |- ...

if metadata.csv has the following format wav_file_name|transcription you can use the ljspeech preprocessor in data/metadata_readers.py, otherwise add your own under the same file.

Make sure that:

• the metadata reader function name is the same as data_name field in session_paths.yaml.
• the metadata file (can be anything) is specified under metadata_path in session_paths.yaml

Change the --config argument based on the configuration of your choice.

python create_training_data.py --config config/session_paths.yaml

This will populate the training data directory (default transformer_tts_data.ljspeech).

python train_aligner.py --config config/session_paths.yaml

First use the aligner model to create the durations dataset

python extract_durations.py --config config/session_paths.yaml

this will add the durations.<session name> as well as the char-wise pitch folders to the training data directory.

python train_tts.py --config config/session_paths.yaml

• Training and model settings can be configured in <model>_config.yaml

• To resume training simply use the same configuration files
• To restart training, delete the weights and/or the logs from the logs folder with the training flag --reset_dir (both) or --reset_logs, --reset_weights

tensorboard --logdir /logs/directory/

You can convert the checkpoint files to hdf5 model weights by running

python checkpoints_to_weights.py --config config/session_paths.yaml

From command line with

python predict_tts.py -t "Please, say something." --config config/session_paths.yaml

Or in a python script

from utils.config_manager import Config
from data.audio import Audio

config_loader = Config(config_path=f'config/session_paths.yaml')
audio = Audio(config_loader.config)
model = config_loader.load_model() # optional: can specify checkpoint name
out = model.predict('Please, say something.')

# Convert spectrogram to wav (with griffin lim)
wav = audio.reconstruct_waveform(out['mel'].numpy().T)

From command line with

python predict_tts.py -t "Please, say something." -c config/session_paths.yaml -w path/to/model_weights.hdf5

Or in a python script

from data.audio import Audio
from model.factory import tts_custom

model, config = tts_custom(config_path='path/to/config.yaml', 
                           weights_path='path/to/weights.hdf5')
audio = Audio(config)
out = model.predict('Please, say something.')

# Convert spectrogram to wav (with griffin lim)
wav = audio.reconstruct_waveform(out['mel'].numpy().T)

• Francesco Cardinale, github: cfrancesco

MelGAN and WaveRNN: data normalization and samples' vocoders are from these repos.

Erogol and the Mozilla TTS team for the lively exchange on the topic.

See LICENSE for details.