TacoTron2 ROS

A ros wrapper for https://github.com/NVIDIA/tacotron2

Setup

• Update submodules

pushd tacotron2_ros/modules/tacotron2
git submodule init; git submodule update
popd

• Install cuda 10.1

pushd
wget https://developer.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.105_418.39_linux.run
chmod +x cuda_10.1.105_418.39_linux.run
sudo ./cuda_10.1.105_418.39_linux.run
# install the toolkit only
popd

• Make a virtual env named .venv after you cd into the project directory and activate the virtual environment. Install python requirements:

virtualenv -p $(which python3) .venv
source .venv/bin/activate
pip install torch torchvision 
            numpy scipy matplotlib 
            librosa==0.6.0 
            tensorflow==1.15 tensorboardX 
            inflect==0.2.5 Unidecode==1.0.22 pyyaml 
            rospkg simpleaudio
pushd tacotron2_ros/modules/apex
export PATH=/usr/local/cuda-10.1/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-10.1/lib64:$LD_LIBRARY_PATH
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
popd

• Download pretrained models

• Download our published Tacotron 2 model: tacotron2_statedict.pt

• Download the WaveGlow published model: waveglow_256channels_ljs_v2.pt

• Copy both into `tacotron2_ros/models/

Usage

rosrun tacotron2_ros tacotron2_node.py

Then publish the text of the message to the topic: /tacotron2/tts

rostopic pub /tacotron2/tts std_msgs/String "Hello human. I am a robot."

Note: strings that are too short generate sound files with long echos and reverberations