This contains code (and libs, if using Windows) necessary to make portable programs for inference with FastSpeech2 (MFA-aligned phonetic) and MB-MelGAN on desktop, along with a simple example. The aforementioned models have to be exported in Tensorflow SavedModel format to be used.

If you want to convert your model to the format that this program expects, you can check out the notebook:

It includes a set of easily understandable and modular classes including a simple English text preprocessor, so you can easily copy and integrate them into your program.

Inference is even easier than in Python.

#include "Voice.h"
// This is a folder name. See Voice.h for information 
Voice LJSpeech("LJ");
std::vector<float> AudioData = LJSpeech.Vocalize("I love see plus plus");
VoxUtil::ExportWAV("voc1.wav", AudioData, 22050);

The demo program is available for download to use for Windows x64 and Linux (Ubuntu 18.04) x64.

To use it, do the following depending on platform:

1. Download the Windows x64 binary and LJSpeech model
2. Download the dependencies
3. Extract both, and place tensorflow.dll in the same folder as the executable you downloaded in Step 1
4. Run

Tested in Kubuntu 18.04 LTS

1. Download the Linux x64 binary and LJSpeech model
2. Extract to whatever directory you like
3. Navigate with terminal
4. LD_LIBRARY_PATH=lib ./TensorflowTTSCppInference

For compiling it yourself, see Compiling below

Compiling the demo depends on what platform. Currently two have been tested:

1. Windows 10 x64; MSVC 2017 (v141)
2. Linux(Ubuntu) x64: GCC 7.5.0

Note that to test out your shiny new build afterwards you'll have to download the LJSpeech model (or make one yourself), it's bundled in any of the above precompiled demo download links.

Due to being too heavy, some dependencies (include + libs) have been excluded from the repo. You can find them here for Windows 64-bit release. Simply put the deps folder into the same place as the solution (.sln) file.

Tested with compiler gcc version 7.5.0 (Ubuntu 7.5.0-3ubuntu1~18.04) . If you've got the same one then you should try my precompiled libs (also includes Tensorflow official C API).

1. Download Linux dependencies and extract the deps folder in the same place as the .pro
2. sudo apt install qt5-default
3. qmake TensorflowTTSCppInference.pro
4. make

For Windows x64 and Ubuntu 18.04 x64 (gcc version 7.5.0), precompiled libraries are included. For other platforms, you'll have to download and compile from source the following dependencies:

1. libPhonetisaurus
2. OpenFST (preferably 1.6.2)

For Tensorflow, you need the C API, which you can grab compiled binaries (or view compile instructions) from here.

The rest (such as CppFlow and AudioFile) are included in the source code.

1. With MSVC compiler, you'll have to use /FORCE as linker command line option, because otherwise the linker throws LNK2005 due to OpenFST linking; the included project is already set up for that. Not sure about GCC, but you'll probably have to do something similar.
2. Tensorflow library malfunctions in debug builds, so only build release.
3. Using MSVC 2017 (v141) compiler in Windows.

• Phonetisaurus (text to phoneme): https://github.com/AdolfVonKleist/Phonetisaurus (ported to Windows)
• OpenFST 1.6.2 (for Windows): https://github.com/kkm000/openfst
• Tensorflow C API: https://www.tensorflow.org/install/lang_c
• CppFlow (TF C API -> C++ wrapper): https://github.com/serizba/cppflow
• AudioFile (for WAV export): https://github.com/adamstark/AudioFile