This repository is a collection of scripts and guidelines that BBC R&D produced while making synthetic voices. We use this implementation of Efficiently Trainable Text-to-Speech System Based on Deep Convolutional Networks with Guided Attention as the foundation for our work. The documentation provided will explain the different steps we took to get the best results out of it, making new voices quickly and cheaply.

The tools included inside dataset were written in Python 3, however, the DC_TTS implementation (forked from Kyubyong's repo) used for training and synthesizing the voice was written in Python 2. This will require you to set up two different Anaconda environments. Just remember to activate the necessary Anaconda environment before you start working with the dataset or synthesis code. You can find more information inside the dataset and the synthesis sections.

We produced a phonetically balanced corpus in English that you can use to create your own voice. Please take a look inside dataset for more information on how to create a training dataset.

For the training, we tune an existing model based on the LJ dataset. You can find more information on how to train a new voice model inside train.

Please take a look inside synthesis for more information on synthesizing speech using our modified script.

We wrote a series of articles detailing our process and technical choices.