LibriVox Forum

I've been listening to Cori Samuel's readings over the last few weeks (specifically: Guide to Men; Frankenstein; War of the Worlds; and Adventures of Mrs Seacole). And my computer has been reading along with me too...

Please accept this is as a small homage : http://reddragon.ai/downloads/cori-rdai_2018-03-22.mp3

Welcome!

That's... umm... creepy. Cool, but creepy. Please don't do that to my voice.

It raises an interesting question: is your voice your own, and are others allowed to synthesize it?

TriciaG :

I imagine the voice ownership question is similar to that in other open source projects : Right now, I'm preparing for a FOSSASIA event I'm doing tomorrow (https://2018.fossasia.org/event/rooms.html#venue-4578) - so understand how it feels to 'put it out there' by releasing content for the world to play with. Clearly, though, we definitely shouldn't be claiming to actually 'be' the person with the original voice that we've trained on.

Hopefully, the real Cori Samuel is flattered by our initial attempts at producing a 'sound-alike'. A couple of things are clear, though : (i) the computer generate voice doesn't actually 'know' anything that real Cori does (so any attempts at impersonation wouldn't stand up for long); and (ii) it can't really discuss current topics very convincingly (because all the speech we have is from older texts). OTOH, I kind of like the idea of personal assistant with such strong views on 'Bachelors' and the Crimea...

All the Best
Martin

Welcome, RedDragonAI

This snippet does sound a lot like Cori. Very interesting. How was it made?

The FBI advises never to say anything to a robo-caller, not Hello, or Quit calling me, certainly do not answer their questions, because their main intent is to get enough of your voice imprint recorded that they can synthesize it. Not sure what they do with that, but it is considered a form of identity theft. Think what they could do with all our audio recordings! Definitely eerie!

Actually, we've been working hard to get the 'song' of the phrasing right - which is one of the things that we like about Cori's readings : She has a clear sense of the shape of the sentences. We're also helped, I guess, by Cori's choice of reading matter, which seems to slant towards first-person 'opinionated' accounts.

The basic process for starting to create a computer voice is outlined in software repos like https://github.com/Kyubyong/dc_tts (which also links to an academic paper and audio samples that they've created). We like to think that we've got some extra ideas, etc, but that's a good resource, if you're interested. Our next steps will probably be to 'clean up' the output/audio processing, since it still has some weird warbling and phase-y effects.

As for the FBI / impersonating issue : Since we've been mostly interested in 'whole sentence' generation (which is, IMHO, something that's needed to take a voice from 'robotic' to something better), trying to replicate small snippets isn't really our domain.

It raises interesting legal issues about peoples privacy which haven't been dealt with - I know the recordings are public domain but does that cover my voice ? . Here people can not use my photo without a clear permission . If my voice is my living can I/should I copyright my voice ?

Anne

Very interesting! But I think I will echo Tricia here:

Please don't do that to my voice.

I've been thinking about this since I heard it, because I have very very mixed feelings about it. It is pretty flattering to provide an inspiration / source of data in this way. And I'd never considered my skew towards first-person 'opinionated' texts, but that's a fair cop!

However, it's also straight from the Uncanny Valley, so rather disconcerting to listen to. And more concerning in terms of its long-term effect on the hobby I love. Once it has learned enough to synthesise words I've never recorded, and the uncanniness has been smoothed out -- where does that leave me? It could be used to read any book the listener wished, whether or not that's something I myself would choose to give my voice to. (I avoid, for example, books that are strongly religious, or contain sexual violence.) And I do sometimes record books commercially -- again, a 'good' version of your software would end that, as I don't have a celebrity brand to encourage people to listen to the original rather than an interpretation. Or, indeed, to fund a legal case to clarify 'ownership' of my voice.

I'd genuinely love to hear the project evolve a voice of its own, whether based on mine, or using other methods. That benefits me -- what's the point of putting hours and hours into my recordings if the difference between me and an AI's version will be relatively minor? -- and it also has potential to benefit people who aren't able to use their own voice (e.g. where they have a disorder of the vocal chords, or are unable to coordinate regular speech.)