We use DeepMind's sparse matmul library to speed-up WaveGRU inference.
$ go get github.com/bazelbuild/bazelisk
$ sudo apt-get install libsndfile1 -y
$ pip install librosa
$ bazelisk build wavegru -c opt --copt=-march=native
$ ./bazel-bin/wavegru --weight ./weight.pickle --mel ./ft.npy --output audio.wav