computer generated text-to-speech

williamjones · Post by **williamjones** » July 24th, 2018, 5:31 pm

Does it comply with the Librivox mission statement for someone to run Gutenberg (or any PD) texts through a Text-to-Speech program and upload them to Librivox? Or do sections have to be recorded by an actual person? I ask this because some sections I have listened to definitely sound computer generated. Is there any way to test this?

Post by **TriciaG** » July 24th, 2018, 5:50 pm

No, it is not LV's mission to use computer generated voices. We use real people, even some that speak somewhat robotically.

msfry · Post by **msfry** » July 24th, 2018, 8:41 pm

But it is a good question. Can a suspicious file be tested to see if it is a text-to-speech generated file? It would be tragic if all our personal efforts could be thwarted by computer generated audio files. Is there anything in our mission statement deterring that?

Post by **annise** » July 24th, 2018, 9:04 pm

The mission statement is definitely people not computers - but we do have to trust our readers. There are couple of readers with an unusual sound but they do breathe.

Anne

williamjones · Post by **williamjones** » July 25th, 2018, 6:31 pm

I am delighted to hear that LV does not condone computer generated files.
Suppose that some files are found whose contents *ARE* convincingly "synthetic".
What would be done about them? And with the person who tried to pawn them off a "real" voice recordings?
Is there a policy for that situation?

-- just curious

Post by **annise** » July 25th, 2018, 6:38 pm

How would we prove it? And what would someone gain by generating them and giving them away ?

Anne

williamjones · Post by **williamjones** » July 25th, 2018, 6:59 pm

annise wrote: ↑
July 25th, 2018, 6:38 pm
How would we prove it? And what would someone gain by generating them and giving them away ?

Anne

I can envision an unfortunate soul of diminished self-worth who would like to brag that he has had over 400 items accepted on LibriVox. "His ego makes him do it!"

Or, more kindly, I can envision a throat cancer survivor who is able to "talk" only with the use of an electronic larynx. These devices work by emitting a square-wave "carrier" tone (a buzz sound) and then modulating it with what's left of any vocal chords. That ever present monotone BUZZ tone is one give away that an electronic device is producing the words.

Lastly, I can envision a reader who is bored silly by the proofing and editing process for their recorded files. If he let's the computer do it, then it will be "word perfect" every time, you see. A lot of meaning of the text normally imparted thru intonation and inflection will be totally lost; the rhyming of poems would be hardly noticeable, but every single word will have been included as written.

A last comment: I asked Mr. Google what free programs or websites were available to produce speech files from text files (Text-to-Speech).
I found 84 of them. And, these were the FREE sources.

I am delighted that a friend turned me on to Librivox. I am spreading the word like crazy among all my friends. Keep up the good work!

k5hsj · Post by **k5hsj** » July 25th, 2018, 10:24 pm

annise wrote: ↑
July 24th, 2018, 9:04 pm
The mission statement is definitely people not computers - but we do have to trust our readers. There are couple of readers with an unusual sound but they do breathe.

Anne

Does LibriVox discriminate against zombies?

ColleenMc · Post by **ColleenMc** » September 1st, 2018, 5:37 am

Not as such, I don't think. It's just that there aren't many books out there that are made up completely of moans and/or the word "braaaaains"...

Colleen

Post by **Availle** » September 1st, 2018, 5:43 am

Well... But then again, PP & Zombies won't be PD for a long time to come. Would make a great DR, I'm sure.

Lushnam · Post by **Lushnam** » January 18th, 2019, 5:40 am

williamjones wrote: ↑
July 25th, 2018, 6:59 pm

annise wrote: ↑
July 25th, 2018, 6:38 pm
How would we prove it? And what would someone gain by generating them and giving them away ?

Anne
… Lastly, I can envision a reader who is bored silly by the proofing and editing process for their recorded files. If he let's the computer do it, then it will be "word perfect" every time, you see. A lot of meaning of the text normally imparted thru intonation and inflection will be totally lost; the rhyming of poems would be hardly noticeable, but every single word will have been included as written.
…

Text-to-speech programs don't always pronounce everything correctly (especially names, but also words with multiple pronunciations like 'read'). Most of them have a very similar rhythm, too (they're not necessarily monotone, though). Unless they have some expensive commercial text-to-speech software, it should be easy to identify which ones are really text-to-speech (by recognizing whichever voice is used). I've personally never heard a LibriVox recording that sounded like text-to-speech to me (and although I've listened to many of them, it's been some few years since I did that). Also, they would probably mispronounce LibriVox. They also don't handle some strange characters in texts very well.

You could prove that recordings are text-to-speech by reproducing the exact same text-to-speech audio files.

I listen to text-to-speech regularly for some things with the voices available on the Kindle Fire HD 8 (6th edition).

lurcherlover · Post by **lurcherlover** » January 18th, 2019, 11:48 am

k5hsj wrote: ↑
July 25th, 2018, 10:24 pm

annise wrote: ↑
July 24th, 2018, 9:04 pm
The mission statement is definitely people not computers - but we do have to trust our readers. There are couple of readers with an unusual sound but they do breathe.

Anne
Does LibriVox discriminate against zombies?

We have a Prime Minister that some have described as a Zombie ... (Not that i necessarily agree).

plaidsicle · Post by **plaidsicle** » January 25th, 2019, 7:05 pm

Lushnam wrote: ↑
January 18th, 2019, 5:40 am
Text-to-speech programs don't always pronounce everything correctly (especially names, but also words with multiple pronunciations like 'read'). Most of them have a very similar rhythm, too (they're not necessarily monotone, though). Unless they have some expensive commercial text-to-speech software, it should be easy to identify which ones are really text-to-speech (by recognizing whichever voice is used).

once upon a time my husband and I tried to listen to a free audiobook of Frankenstein that he'd found on Spotify, and it was obviously recorded by a machine-- the cadence and timing were all wonky and it was insufferable to listen to. I told husband that we should've gone to LV first! (and we did ultimately listen to one of our versions)

as DPL for some poetry a while back, I once had one reader on the project raise suspicions about whether another's recording was human or not. I'm sure it will get harder and harder to always be able to tell. it is good to know that there is at least some method of proving it. but does anyone at LV really have the time to vet suspicious recordings and find proof by trying to replicate them? that sounds impractical.