As a follow-up on Open-source vocal cloning (speech-to-speech neural style transfer), I want to create a voice clone. Unfortunately, the answers in the thread above do not apply to my language and dialect (European Portuguese). So I would like to DIY a voice clone with a text-to-speech service (Google, Polly, Azure) and a neural network trained to transfer the style. The difference with most vocal cloning systems is that I can generate any amount of aligned data: the system and I both read the same script.
A quick search for
vocal cloning with parallel|aligned data yields no results specific to this problem. But from reading a few papers and their emphasis on non-aligned data (e.g., "learned in an unsupervised way from unaligned data" here), I get the impression that style transfer with aligned data is a simpler problem that may have established solutions.
What papers, code, or architecture implement vocal cloning with aligned data?