speech synthesis c++ inference for voicenet

ttsflow can be built with both google's bazel build tool and gnu's make. We provide compatibility with bazel for convenience to build from scratch once cloned. However, the final executable file may be really big, which is nearly 60M. The reason is that many unused tensorflow codes are compiled and linked into the final executable file staticly. Maybe it is more suitable to build with gnu's makefile, especially when you want to incorporate tensorflow into your project, so we provide makefile too.

It's quite simple to build this project with bazel, just type:

cd ./third_party/tensorflow
./configure
cd ../../
bazel build -c opt --config=opt ttsflow/examples/...

The generated executable file is located in bazel-bin/ttsflow/examples/ttsflow_test

In order to build with tensorflow, you should first compile tensorflow as dynamic-link library.

To build libtensorflow_cc.so for c++, just run:

cd ./third_party/tensorflow
./configure
bazel build -c opt --config=opt //tensorflow:libtensorflow_cc.so

Another prerequisite is protobuf, which should be installed in ttsflow/third_party/protobuf:

cd ./third_party/protobuf
./autogen.sh
./configure --prefix=`pwd`
make
make install

As for vocoder, we use the world vocoder, which is placed in ttsflow/third_party/World:

cd ./third_party/World
make

To compile this project, just directly type make, and a symbolic link named ttsflow_main will be created under the root dir.

make

# Run ttsflow
./ttsflow_main

Also, you can run the test cases after compiling the test files.

make test