Awesome

Pytorch Implementation of DC-TTS for Emotional TTS

This fork is modified to work for transfer learning for low-resource emotional TTS, as described here.

Training

1. Install the dependencies using pip install -r requirements.txt
2. Preprocess the EmoV-DB dataset using process_emovdb.py
3. Change the logdir argument in hyperparams.py. Other parameters can be edits optionally. DO NOT edit these hyperparameters.
4. Add the path to the pre-trained Text2Mel model in the logdir
5. Comment this line if you are not running the train-text2mel.py file for the first time.
6. Run the training script like - python train-text2mel.py --dataset=emovdb

Synthesis

1. Write the sentences that you want to generate here
2. Add the checkpoint for the fine-tuned Text2Mel model in place of this line
3. Edit the paths for the output.
4. Run the synthesis script like - python synthesize.py -- dataset=emovdb

Readme of the original repository

PyTorch implementation of Efficiently Trainable Text-to-Speech System Based on Deep Convolutional Networks with Guided Attention based partially on the following projects:

• https://github.com/Kyubyong/dc_tts (audio pre processing)
• https://github.com/r9y9/deepvoice3_pytorch (data loader sampler)

Online Text-To-Speech Demo

The following notebooks are executable on https://colab.research.google.com :

• Mongolian Male Voice TTS Demo
• English Female Voice TTS Demo (LJ-Speech)

For audio samples and pretrained models, visit the above notebook links.

Training/Synthesizing English Text-To-Speech

The English TTS uses the LJ-Speech dataset.

1. Download the dataset: python dl_and_preprop_dataset.py --dataset=ljspeech
2. Train the Text2Mel model: python train-text2mel.py --dataset=ljspeech
3. Train the SSRN model: python train-ssrn.py --dataset=ljspeech
4. Synthesize sentences: python synthesize.py --dataset=ljspeech
   • The WAV files are saved in the samples folder.

Training/Synthesizing Mongolian Text-To-Speech

The Mongolian text-to-speech uses 5 hours audio from the Mongolian Bible.

1. Download the dataset: python dl_and_preprop_dataset.py --dataset=mbspeech
2. Train the Text2Mel model: python train-text2mel.py --dataset=mbspeech
3. Train the SSRN model: python train-ssrn.py --dataset=mbspeech
4. Synthesize sentences: python synthesize.py --dataset=mbspeech
   • The WAV files are saved in the samples folder.