structure and thoughts.txt

WaveRNN seems competetive with latest end-to-end wave nets

TTS Model needs to:
* Be trained trained on multiple different languages and be able to extract between them
* Have a smooth latent space for the speaker (maybe a latent space for language and speaker combination?)

Translation model needs to:
* Be trained on the phenomes being about equally long for input and answer

Is there a way to merge the STT, translation and TTS??

Voice Conversion seems like a way to directly transform a voice to another language. 
Currently it's used to transform one speaker into another one, but we can look to extract a language independant representation (rather than speaker independant)
and project that into another language, keeping the specifics of the speaker.

Contrastive loss can be used to make sure the latent space is smooth

Must disentangle speaker identity from natural and language/accent based prosody