Skip to content

Desktop application for neural speech synthesis written in C++

License

Notifications You must be signed in to change notification settings

ZDisket/TensorVox

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

TensorVox

TensorVox is an application designed to enable user-friendly and lightweight neural speech synthesis in the desktop, aimed at increasing accessibility to such technology.

Powered mainly by TensorFlowTTS and also by Coqui-TTS and VITS, it is written in pure C++/Qt, using the Tensorflow C API for interacting with Tensorflow models (first two), and LibTorch for PyTorch ones. This way, we can perform inference without having to install gigabytes worth of Python libraries, just a few DLLs.

Interface with Tac2 model loaded

Try it out

Detailed guide in Google Docs

Grab a copy from the releases, extract the .zip and check the Google Drive folder for models and installation instructions

If you're interested in using your own model, first you need to train then export it.

Supported architectures

TensorVox supports models from three repos:

  • TensorFlowTTS: FastSpeech2, Tacotron2, both char and phoneme based and Multi-Band MelGAN. Here's a Colab notebook demonstrating how to export the LJSpeech pretrained, char-based Tacotron2 model:
  • Coqui-TTS: Tacotron2 (phoneme-based IPA) and Multi-Band MelGAN, after converting from PyTorch to Tensorflow. Here's a notebook showing how to export the LJSpeech DDC model:
  • jaywalnut310/VITS: VITS, which is a fully E2E model. (Stressed IPA as phonemes) Export notebook:

Those two examples should provide you with enough guidance to understand what is needed. If you're looking to train a model specifically for this purpose then I recommend TensorFlowTTS, as it is the one with the best support, and VITS, as it's the closest thing to perfect As for languages, out-of-the-box support is provided for English (Coqui and TFTTS, VITS), German and Spanish (only TensorFlowTTS); that is, you won't have to do anything. You can add languages without modifying code, as long as the phoneme set are IPA (stressed or nonstressed), ARPA, or GlobalPhone, (open an issue and I'll explain it to you)

Build instructions

Currently, only Windows 10 x64 (although I've heard reports of it running on 8.1) is supported.

Requirements:

  1. Qt Creator
  2. MSVC 2017 (v141) compiler

Primed build (with all provided libraries):

  1. Download precompiled binary dependencies and includes
  2. Unzip it so that the deps folder is in the same place as the .pro and main source files.
  3. Open the project with Qt Creator, add your compiler and compile

Note that to try your shiny new executable you'll need to download a release of program as described above and replace the executable in that release with your new one, so you have all the DLLs in place.

TODO: Add instructions for compile from scratch.

Externals (and thanks)

Contact

You can open an issue here or join the Discord server and discuss/ask anything there

For media/licensing/any other formal stuff inquiries, send to this email: 9yba9c1y@anonaddy.me

Note about licensing

This program itself is MIT licensed, but for the models you use, their license terms apply. For example, if you're in Vietnam and using TensorFlowTTS models, you'll have to check here for some details