Moho Forum

Here is a video of a test of the text to speech from:

https://wellsaidlabs.com/

No animation, just a test of the text to speech generator.



Dale

That's amazingly good for a Text-to-Speech program.

I haven't had a chance to play with the WellSaid trial yet. How much are you able to tweak the performance?

Greenlaw wrote: ↑Wed Feb 17, 2021 1:28 am That's amazingly good for a Text-to-Speech program.

I haven't had a chance to play with the WellSaid trial yet. How much are you able to tweak the performance?

Yes, There are the words I tweak in the reading.

chamois -edit to- sham-moss

Jabone -edit to- Jab-bone

Sutra -edit to- Su-tra

Tatana -edit to- Tat-too-na

As you can see mostly names, But simple words can hang it up because the spelling and normal pronunciation don't match.
If you go to the very end of the video the lines are:
-------
“Yes.”
“It's is ‘yes, sister’ from now on.”
-----
it sounds like:
"Yes-itsyes sister from now on."

I think it is because there should be a pause longer than the generator gives for the text.

There is an output for each line so you can go in and fix a lot. I only wish that it had inflection like written French.

I have sent in a request for slight speed, pitch, and a 8 band equalizer. Also child voices.

They wrote me back that sent it along and another 15 voices are coming.

Dale

It would be cool to be able to set emotions like surprise, sadness etc for specific lines.

The reason why this software works:

https://www.google.com/amp/s/techcrunch ... umans/amp/

Dale

Good article! The observation of how inconsistencies make a human voices sound natural is very apparent here. I wondered if deep learning was involved, and it's interesting to read how it was used to model the patterns.

Thanks for sharing that link!

Greenlaw wrote: ↑Sun Feb 21, 2021 6:48 pm Good article! The observation of how inconsistencies make a human voices sound natural is very apparent here. I wondered if deep learning was involved, and it's interesting to read how it was used to model the patterns.

Thanks for sharing that link!

Here is an ugly thought. If all it takes is two to four hours of someone speaking to simulate another person's voice, then how hard would it be to replace their mouth and voice?

With Aftereffects, Moho, and WellSaid I could get any politician saying anything in a video. Trump says, "...." Or Biden said, "...." and it would be very hard to tell it has been doctored.

Dale

The tech is already with us to fake it.

Deepfakes and Virtual Surrogates are the Future !!!

My intentions for this technology tend to be frivolous and not especially sinister.

(Of course, that's exactly what he would say, isn't it.)

Just for fun - I was playing with the demo voices at: https://cloud.google.com/text-to-speech#section-2

Some voices sound a little robotic, but some (like: English US / en-US-Wavenet-I) are OK.

It was also quite interesting to put english text onto non-english speakers - like "Arabic, France, Italiano" - where 'they' still speak english but with an accent.

p.s. You may need to select different WaveNet - Voice Name settings to get decent results.

Amazing quality. I have always had to patch my narration track together because I can't finish a sentence without stumbling.