Parallelizing VOCA to Reduce the Latency

VOCA (Voice Operated Character Animation) is an end-to-end system that generates 3D facial animations. In this work, we parallelized VOCA to reduce its latency. In this research, we made three design decisions : data parallelism over task parallelism for simplicity, multiprocessing over multithreading for high performance , and shared memory over message copying to reduce communication overhead. We evaluated the performance of our proposed method with a various dataset on different Google Cloud virtual machine types. With our parallelization method, we showed a performance close to the theoretical speed-up with an audio of 1 hour length where the latency is reduced from 3453 sec to 587 src using 6 processors. Our parallelization method works well especially when the input text and speech length are long which occurs frequently in the real-world settings.