playing around with text-to-speech
Aug. 13th, 2024 01:09 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
queenluawhenever i'm editing a piece that i'm being somewhat-to-very tryhard about, i usually make an effort to read the piece aloud to myself. ideally i'd read the whole thing, slowly, audiobook-style, but more often i'm doing some mix of that + "just muttering passages quietly to myself." it's pretty good for catching the sorts of errors that the brain's too good at "filtering out" while reading (e.g. repeating a word, an awkward dialogue tag, etc).
but, i got curious the other night about the state of text-to-speech software, because hey, that's one of the few domains where "just throw more GPUs at it" does seem pretty useful, and i ran out of podcasts for this week's commute, and yeah i'm absolutely vain enough to make a computer audiobookify my own shit haha.
so, lo, here's the random software i decided to play with after a google search. cursory observations:
* these voices are pretty good. like, you're not going to mistake them for a human reader (in particular there's a weirdly "clipped" quality to the way they finish a lot of their sentences, and a sort of monotony/regularity to the way in which they do end-of-sentence/end-of-paragraph-type pauses that sounds distinctly unnatural when you listen for longer periods of time), and i'd certainly rather pay money for the human-read audiobook version of any narrative i actually wanted to enjoy (the lack of any attempt to create different "voices" for each character is a huge drawback), BUT, this is leagues better than the standard-accessibility-suite robo-voice i remember from 00s-era mac osx lol. reasonably pleasant, not too grating, totally works for "being forced to hear my own writing" purposes
* this particular software absolutely cannot handle italics, rip. admittedly this ends up serving as a good reminder that i should be using italics less anyway, but, y'know, sometimes i do need that extra emphasis!!!
* the "audiobook" is excellent for forcing me to notice "stupid" errors (repeated words etc), and i think it miiiight give me a better sense of pacing fuckups? in the sense that, if i've been staring at a wall of text for a while, it's hard for me to get a sense of where a reader might lose interest, whereas if these words are washing over me while i'm in some standstill traffic on the f$@&*ing bridge again, i'm getting a decent intuitive sense of "ok how long is this part going on for & is it actually interesting"
* (unfortunately if i'm listening while in some standstill traffic on the f$@&*ing bridge again, i can't exactly, uh, stop to take notes or fix the manuscript right there, so i'm relying on "just remembering what sounded off," but eh a little mental exercise is good for you)
* i certainly wouldn't want to use this as the only source of reading-aloud-ness since the computer-voice-guy makes some repeated "flow" choices that i just think are WRONG lol. for instance, i think the voice guy gives literally every comma about equal weight, which makes any standalone super-short demarked-by-a-comma phrase, like this one, sound REALLY awkward in a way that i think any ordinary human reading a passage will not find awkward.
* tellius inside baseball observations: this tool pronounces tibarn as "TYE-barn", pronounces as reyson "rey-SON", and pronounces naesala "nae-SA-la," all of which i deem the WRONG way to pronounce those respective names lol. (i'm aware FE Heroes disagrees with me re: naesala, but that just means heroes is wrong too sorry!!!) also nikolias gets pronounced "neh-COE-li-as" which is ALSO wrong. and i made that one up so i'm objectively right here for sure lmao
further observations to be reported when/if they prove interesting
but, i got curious the other night about the state of text-to-speech software, because hey, that's one of the few domains where "just throw more GPUs at it" does seem pretty useful, and i ran out of podcasts for this week's commute, and yeah i'm absolutely vain enough to make a computer audiobookify my own shit haha.
so, lo, here's the random software i decided to play with after a google search. cursory observations:
* these voices are pretty good. like, you're not going to mistake them for a human reader (in particular there's a weirdly "clipped" quality to the way they finish a lot of their sentences, and a sort of monotony/regularity to the way in which they do end-of-sentence/end-of-paragraph-type pauses that sounds distinctly unnatural when you listen for longer periods of time), and i'd certainly rather pay money for the human-read audiobook version of any narrative i actually wanted to enjoy (the lack of any attempt to create different "voices" for each character is a huge drawback), BUT, this is leagues better than the standard-accessibility-suite robo-voice i remember from 00s-era mac osx lol. reasonably pleasant, not too grating, totally works for "being forced to hear my own writing" purposes
* this particular software absolutely cannot handle italics, rip. admittedly this ends up serving as a good reminder that i should be using italics less anyway, but, y'know, sometimes i do need that extra emphasis!!!
* the "audiobook" is excellent for forcing me to notice "stupid" errors (repeated words etc), and i think it miiiight give me a better sense of pacing fuckups? in the sense that, if i've been staring at a wall of text for a while, it's hard for me to get a sense of where a reader might lose interest, whereas if these words are washing over me while i'm in some standstill traffic on the f$@&*ing bridge again, i'm getting a decent intuitive sense of "ok how long is this part going on for & is it actually interesting"
* (unfortunately if i'm listening while in some standstill traffic on the f$@&*ing bridge again, i can't exactly, uh, stop to take notes or fix the manuscript right there, so i'm relying on "just remembering what sounded off," but eh a little mental exercise is good for you)
* i certainly wouldn't want to use this as the only source of reading-aloud-ness since the computer-voice-guy makes some repeated "flow" choices that i just think are WRONG lol. for instance, i think the voice guy gives literally every comma about equal weight, which makes any standalone super-short demarked-by-a-comma phrase, like this one, sound REALLY awkward in a way that i think any ordinary human reading a passage will not find awkward.
* tellius inside baseball observations: this tool pronounces tibarn as "TYE-barn", pronounces as reyson "rey-SON", and pronounces naesala "nae-SA-la," all of which i deem the WRONG way to pronounce those respective names lol. (i'm aware FE Heroes disagrees with me re: naesala, but that just means heroes is wrong too sorry!!!) also nikolias gets pronounced "neh-COE-li-as" which is ALSO wrong. and i made that one up so i'm objectively right here for sure lmao
further observations to be reported when/if they prove interesting
no subject
Date: 2024-08-13 08:39 pm (UTC)super fascinating!!! having used dragon naturallyspeaking once or twice back in the 00's as you mentioned it's awesome hearing that it's improved since then. (tbh automatic captions/translations are another area i've seen a distinct marked improvement over the decades - i recently participated in a meeting with only mexican-spanish speakers and understood 95% of it right off of pure captioning, it was awesome, 10/10 would be a fly on the wall again.
cheering for u in the home stretch there!
no subject
Date: 2024-08-13 11:42 pm (UTC)