Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model

Re:Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model by dharma1 on Thursday March 30, 2017 @05:09AM (#13993919)

It's not really style transfer, but for a new speaker model, you just need to train each speaker with a dataset of 25 hours audio with time matched accurate transcriptions.

In the case of David Attenborough, I'm pretty sure that amount of data is available with subtitled BBC content.

[ Parent | on HN ]

Re:Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model by dnautics on Thursday March 30, 2017 @12:25PM (#13996754)

Oh I doubt that will be necessary. I bet you could find a useful "feature vector" by repeating the process on about 10-20 new voices - then extract a set of words (let's pessimistically say 100) that activate these features the most - then have a person retrain with 5 examples of each of these 100 words (an hour, at the most).

I honestly wouldn't be surprised if you could find a really feature rich sentence (say 5 words long) that you could use to crack pretty much all voice-activated biometric password systems.

[ Parent | on HN ]

Re:Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model by slinger on Thursday March 30, 2017 @04:41PM (#13998683)

Is voice-activated password a thing? I mean, I only saw that in sci-fi movies and never in real life.

[ Parent | on HN ]

Re:Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model by dharma1 on Thursday March 30, 2017 @05:27PM (#13999021)

Yes, voice biometrics are used in telephone banking for instance

[ Parent | on HN ]

Re:Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model by dharma1 on Thursday March 30, 2017 @05:32PM (#13999068)

yep, I'm sure there would be a lot of similarity between models that could be used to reduce training time

[ Parent | on HN ]