This repository contains the Idiap Text-to-Speech system developed at the Idiap Research Institute, Martigny, Switzerland.
Contact: bastian.schnell@idiap.ch

It is an almost purely Python-based modular toolbox for building Deep Neural Network models (using PyTorch) for statistical parametric speech synthesis. It provides scripts for feature extraction and preparation which are based on third-party tools (e.g. Festival). It uses the WORLD vocoder (i.e. its Python wrapper) for waveform generation. The framework was highly inspired by Merlin and reuses some of its data preparation functionalities. In contrast to Merlin it is intended to be more modular and allowing prototyping purely in Python.

It comes with recipes in the spirit of Kaldi located in separate repositories.

IdiapTTS is distributed under the MIT license, allowing unrestricted commercial and non-commercial use.

IdiapTTS is tested with: Python 3.6

Follow the instructions given in INSTALL.md.

Instructions to run specific experiments are in the README files of the respective egs repositories:

• https://github.com/idiap/idiaptts_egs_ljspeech /s1 contains TTS with a duration and acoustic model.

Instructions to produce results similar to those reported in the paper can be found at https://github.com/idiap/idiaptts_egs_blizzard08_roger s1/.

Instructions to reproduce the results of the paper can be found in https://github.com/idiap/idiaptts_egs_blizzard08_roger s2/.

Instructions to produce results similar to those in the paper can be found at https://github.com/idiap/idiaptts_egs_vctk s1/.