Using Amazon Polly Text to Speech to Mimic the Intonation of a Real Speech Waveform from an Audio File

Question

I want to produce very realistic sounding speech using Amazon Polly -- more realistic than the results produced by allowing Amazon Polly to generate speech audio output from normal text input. In order to do this, I want to tell Polly to do its best to exactly match a given sample of real speech audio file input that was voiced/recorded by a voice actor. In other words, I want to convert the voiced speech produced by the voice actor into a sequence of some kind of codes (eg phonemes, timing codes, etc) suitable as input to the polly api, which can then produce an audio file output that should mimic the original recorded input as closely as possible.

Does some AWS api take speech audio file as input, and turn it into a series of codes (eg phonemes and timing info) that can then be fed into the polly api to generate corresponding speech back? If not, is there some other api that would be appropriate to be used for the first step?

Why? Using Polly to create speech the 'normal way' -- by feeding text into the polly api -- tends to produce not sufficiently natural sounding speech (though it may be the best text to speech engine currently available). The Polly API speech produced the 'normal way' does not have the correct timing/prosody. We've tried to manually convert speech word-by-word into phonemes and manually add timing information, and then feed this into the Polly API as input, but this is both very laborious, and doesn't produce a very satisfactory result. We have used this in apps. For particularly important parts of the speech, we need to get better quality. We can't use recorded audio from a voice actor, as one might in other circumstances, because then it won't match the rest of the speech, produced by the Polly API.

BTW, before you vote this post down because it doesn't include a little code snippet... consider that this is still a coding problem, just not in the form of 'help me edit my code snippet to do x' -- though a code snippet answer would be a fine form for an answer to the question if such an answer exists! This is actually a broadly important and general coding problem that potentially affects any coder needing to create realistic speech through code.

Thanks

Answer 1

Does some AWS api take speech audio file as input, and turn it into a series of codes (eg phonemes and timing info) that can then be fed into the polly api to generate corresponding speech back?

Do you seriously think that Amazon created some secret API hidden it from the users but some brave hacker will disclose you undocumented API access code here?

If not, is there some other api that would be appropriate to be used for the first step?

This exists in open source, unrelated to AWS of course

https://github.com/KinglittleQ/GST-Tacotron