Generating accents of the English language using WaveNet: a formant analysis study

Abstract

WaveNet is a neural network that trains on and generates raw audio waveforms. It does this with an exceptional, almost human-like performance, especially in the text-to-speech-area. This paper investigates the capabilities of WaveNet concerning accents. We discuss two ways to measure the similarity between the generated accent and the accent from the VCTK Corpus. The experiments could not be conducted because of the absence of a working implementation of local conditioning. There was no significant difference between the results of the different learning rates.