Not really happy with a lot of the content available and decided I was interested in trying to produce my own. I'm not a girl however, and I'm not really interested in paying a girl to read for me, and I figure most people aren't here either, so I looked into TTS options. B4mb1 and 54m4nth4 B4nd|3r are TTS and are really effective, so spreading good TTS info might get the ball rolling on more community produced content, found a lot of info that I'm gonna share.
Basically, existing TTS options are voices that can be installed to your system and played through various TTS readers, like Text Aloud (paid) and Balabolka (free). I've gone through a ton of the voices available out there and the only ones that are any good really are IVONA's Salli and Neospeech's Julie.
https://thepiratebay.org/torrent/8648479/IVONA_Text_to_Speech_with_crack_(All_voices)
https://thepiratebay.org/torrent/5351567/Neospeech_Julie_-_Works_w__MS_Office_2007-10_Best_Text_to_Speech
AT&T Crystal is next best, but it sounds kind of drowsy and artificial. Drowsy might be what someone is going for though.
Google Translate is okay, natural reading, mildly stilted, kinda disinterested.
This site lets you type in stuff and have it read.
https://www.naturalreaders.com/online/index.html
Amanda and Sharon are pretty okay.
Selene and Audrey are slightly less okay, but still okay. Obviously nothing really compares to Salli and Julie, but it's worth bringing up options, for variety. Hell, some actual hypnotists have worse voices than these.
Figuring out how to trick the existing TTS programs into giving words the correct inflection is probably the route to go in making them feel less artificial. It seems like there's different inflections added in text aloud based on how the sentence is written, like saying the last word of a sentence at a higher pitch.
Looking into this more, seems Google has some Deep Learning TTS application called Wavenet, which they published a paper on, but did not release their version to the public.
https://deepmind.com/blog/wavenet-generative-model-raw-audio/
https://arxiv.org/pdf/1609.03499.pdf
Then there's Merlin, which is another Neural Net, which works through text reader front-ends like Festival (different file format than the ones above)
https://github.com/CSTR-Edinburgh/merlin
Documentation and guide to building your own voice
https://cstr-edinburgh.github.io/merlin/
http://neospeech.com/
https://www.ivona.com/us/about-us/voice-portfolio/
The basis of Neural Net TTS applications is they take in a ton of vocal samples, of a person and train the neural network on these samples to produce the ability to predict what the next wavepattern will be, based on a text input. So they tell the AI a text input, it guesses what the next wavepattern is, the AI's output is compared to the actual next wavepattern from the vocal sample, and if it's inaccurate, then the inaccuracies are backpropagated through the nodes of the neural net to calibrate it to produce a more accurate output next time. It does this across a wide variety of text samples, with matching vocal samples, until it can accurately predict what the vocal sample sounds like from just the text, and then accurately predict what voice samples would sound like from text that there is no vocal sample for, effectively being able to perform text to speech.
Public implementation of Google's Wavenet on Tensorflow, which apparently works really well.
https://github.com/ibab/tensorflow-wavenet
Fast wavenet? Supposed to be fast. Cuts out a lot of unnecessary steps according to the description.
https://github.com/tomlepaine/fast-wavenet
Keras implementation of wavenet (slower)
https://github.com/basveeling/wavenet
This has information on neural nets, and speculation on their use for generating music
http://deepsound.io/wavenet_first_try.html
Wavenet apparently requires a good GPU (go figure). Also takes a lot of time to process samples.
This is an open source TTS engine, with its own tools for creating new voices with it, Merlin is designed to work with this as the front-end.
http://festvox.org/festival/
Manual for building a new voice is in here:
http://festvox.org/docs/manual-2.4.0/festival_toc.html
Given we have a ton of hypnotist files without background audio sitting around, it might be possible for us to build some custom voices based on existing hypnotists, and use them to create new files.
I'm a bit too busy to undertake a project like this right now, but I'm reposting the info in case anyone else wants to try picking it up, or is interested in just messing around with Julie and Salli.
If anyone knows anything else, please share.