Adaptive Speech Synthesis and training using low-cost AI Computing

摘要

Computer synthesis of human speech has long been possible. It is inexpensive to build the infrastructure and computing to apply subtle inflections, vocal character, phrasing, and modulation; however, these additions come at a price - not the expense of the hardware, but the large number of constants to be set and time-varying parameters that must be dynamically controlled. Although these parameters can be set manually, it is not time efficient, and having one or even dozens of dynamic descriptors is not enough to emulate realistic speech. When the vocal representation is converted to a 2-D image, an adaptation of the spectrogram, it can be easily analyzed using a CNN (convolutional neural network). Existing low-cost tools for CNN analysis can now be applied. This structure, coupled with a fully-connected neural network, can serve as a human-interactive interface for real-time training of the AI synthesizer, emulating phrasing and even accents of the trainer. This paper details the structure of the synthesizer and the training AI, and presents some practical, functional examples exploring this technique using KERAS, LabVIEW and MATLAB.