Development of a Neural Network Library for Resource Constrained Speech Synthesis

Machine learning frameworks, like Tensorflow and PyTorch, use GPU hardware acceleration to deliver the needed performance. Since GPUs require a lot of power (and space) to operate, typical use cases involve high-performance servers, with the final deployment available as a cloud service. To address limitations of this approach, AI Accelerators have been proposed. In this context, we have designed and implemented a library of neural network algorithms, to efficiently run on “edge devices”, with AI Accelerators. Moreover, a unified interface has been provided, to allow easy experimentation with various neural networks applied to the same dataset. Here, let us stress that we do not propose new algorithms, but port known ones to, resource restricted, edge devices. The context is provided by a speech synthesis application for edge devices that is deployed on an NVIDIA Jetson Nano. This application is to be used by social robots for real-time off-cloud text-to-speech processing.