Released Models

English

TTS Models	Dataset	Commit	Audio Sample	Details
Tacotron2	LJSpeech	branch	---	details
Tacotron2 with Graves (soon)	LJSpeech	branch	---	details
Tacotron2 DDC	LJSpeech	72a6ac5	voice samples	Trained with DDC and includes PyTorch, Tensorflow and TFLite models. Check Colab notebooks or notebooks folder.
Glow-TTS	LJSpeech	08394e4	---

Speaker Encoder Models	Dataset	Commit
Speaker-Encoder-iter25k	LibriSpeech	...

Vocoder Models	Dataset	Commit	Details
ParallelWaveGAN	LJSpeech	72a6ac5	Trained using TTS.vocoder. It produces better results than MelGAN model but it is slightly slower. Check notebooks for testing.
Multi-Band MelGAN	LJSpeech	72a6ac5	Trained using TTS.vocoder. It is the fastest vocoder model. Check notebooks for testing.
WaveRNN models			go to repo for the models. (Soon to be deprecated)
Full-Band MelGAN	LibriTTS	c514628	Trained using TTS.vocoder. Generic vocoder that can sample any voice. Sampling rate 24Khz. To use with a different sampling rate follow this issue.

Simple packaging - self contained package that runs an HTTP API for a pre-trained TTS model

How to use:

1. Create a fresh virtual environment with Python 3.6
2. $ apt-get install espeak libsndfile1
3. $ pip install python_package_url_from_table_below
4. $ python -m TTS.server.server
5. Open http://localhost:5002

Model	Dataset	Python package	nginx/uWSGI config files
Tacotron 2 + Forward Attention + PWGAN	LJSpeech	TTS-0.0.1+92aea2a-py3-none-any.whl	tts-nginx-uwsgi.zip

The server is a Flask application. For deployment with multiple workers see the nginx/uWSGI config files also linked in the table above. Pass --use_cuda 1 to use GPUs when available.

German

Soon...

Portuguese

model details by @Edresson