Wikipedia talk:WikiProject Spoken Wikipedia

Main Page | Creating Recordings | Spoken Articles | Talk | Being Recorded | Review | Participants

---

Bulk removals?

Any comments on these bulk removals? "rv text-to-speech added without consensus" and many more at Opencooper (talk · contribs).

Is the project still live? Are they all to be removed now? Where was this decided?

I don't understand the "added without consensus" comment either. Surely they're covered by the general project and don't need per-article specific permissions?

@Alanasings: Andy Dingley (talk) 00:05, 9 February 2024 (UTC)

I believe this is related to this discussion last December: Wikipedia talk:WikiProject Spoken Wikipedia/Archive 12 § Articles being read out by AI voices/screen readers? isaacl (talk) 00:16, 9 February 2024 (UTC)

Perhaps my edit summary wasn't clear enough. I wasn't removing regular spoken articles read by humans, but those created by an AI text-to-speech (TTS). These were all uploaded by a single user, User:Alanasings, without gaining consensus for adding them first. I was planning to write a post on their talk page after finishing, but didn't get to removing all of them.

As for the "why", these spoken articles are meant to be created by humans. TTS technology has been around forever. If that was perfect and could easily replace humans, this project wouldn't exist, and those would be automatically generated for articles.

The quality of these TTS uploads is, frankly, crap. Firstly, it's clear that the creator isn't reviewing or editing these. Blue Book (magazine)'s TTS has 30 seconds of silence and starts with "test, test, test", and Eileen (novel) has 10 seconds of silence. Over at Evika Siliņa and Zach Galifianakis, it didn't even say the names of the subjects correctly. Same at Juanpa Zurita, where it also read out the infobox. At womance, it couldn't figure out how to pronounce the title, which is meant to rhyme with "romance". At Link rot, it says "reference rah" (not "rot"). At Black Mask (magazine), it can't say "subgenre". Deals Gap, North Carolina has bad enunciation. At Sceloporus malachiticus, it says "6 8 inches" instead to "6 to 8" for "6–8", and omits the conversion in parenthesis. True Confessions (magazine) has the title said twice. Every single one of these has unnatural cadence.

I could go on for days on how all of these TTS recordings are subpar, but it's evident that they do readers a disservice compared to a human reader taking the time to pronounce things correctly, respect punctuation and natural pauses, and actually understand what is being said. The uploader should have never flooded Wikipedia with these without consulting the community. Opencooper (talk) 01:01, 9 February 2024 (UTC)

I don't think the files are "sub par." A lot of work went into them. Alanasings (talk) 13:59, 10 February 2024 (UTC)

If you don't think they're subpar, can you address the unacceptable quality of the pronunciation and cadence compared to a human reader? For many of these, it didn't even say the subject's name correctly, which is the first basic thing that a human would have made sure to get right. You're clearly not reviewing them since you published them regardless of so many glaring errors. The amount of time spent creating them doesn't excuse their fundamental shortcomings and editorial inappropriateness. Clicking a button to make AI do something for you doesn't absolve you of all responsibility for the result. Opencooper (talk) 20:12, 10 February 2024 (UTC)

Some of them might have mispronunciations. I don't think there is any problem with the cadence, especially in the latter ones. There is no "editorial inappropriateness." They were reviewed. There is more it to than "pushing a button." You seem to have a very strong opinion abut this that borders on rude. I don't care if you take them down or not. I am not interested arguing with you on Wikipedia. I won't be making any more of the spoken files. By the way, this is how Wikipedia loses enthusiastic editors. Alanasings (talk) 20:33, 10 February 2024 (UTC)

I, too, would rather articles have text-to-speech recordings than none at all. I appreciate User:Alanasings' work here and hope to understand how these recordings are problematic. Perhaps there's a way to clarify that the recordings are text-to-speech?, and/or, perhaps these can be kept until a human version is recorded? ---Another Believer _(Talk) 19:40, 19 February 2024 (UTC)

Readers content with automated speech versions can use screen readers, and thus benefit from being able to hear the latest version (or any earlier version, if they wish). isaacl (talk) 23:30, 19 February 2024 (UTC)

Agreed. Though I feel @Opencooper should have opened a discussion before mass-deleting, I don't see how computer-generated voice recordings are an improvement over screen readers. This aim of this project, as far as I can tell, is to provide recordings of articles read out by humans, not AIs. Funcrunch (talk) 03:27, 20 February 2024 (UTC)

@Another Believer: I've already pointed out fundamental issues with these recordings, including not even pronouncing the subject correctly, so these are not better than nothing. Every single recording when put under a little scrutiny has problems. Picking randomly, at Elisabeth Boehm, it doesn't say "Tatar" and "Hungarian" correctly, and even forgets how to speak English when saying "Governorate".

Keeping these in articles will actually prevent human versions from being recorded, since people will think the article has already been covered, and they'll easily be outpaced by low-quality AI content generated by the hundreds.

@Funcrunch: I know you mostly agree with me, but it's curious how the onus is on me, and you don't feel Alanasings should have opened a discussion before flooding the project with these. They themselves were fine with their removal, and I waited a week and there were no objections, so I proceeded. I should have included a pointer to this discussion in the summaries though. Opencooper (talk) 04:33, 20 February 2024 (UTC)

@Opencooper From their response upthread, I wouldn't agree that @Alanasings was "fine" with the removals, but I don't want to press the issue. Funcrunch (talk) 05:13, 20 February 2024 (UTC)

No, I wasn't "fine" with it. Alanasings (talk) 00:31, 15 March 2024 (UTC)

I linked earlier to the related discussion in December. isaacl (talk) 04:35, 20 February 2024 (UTC)

Coming in a bit late, but is there a consensus on whether or not to remove the AI screenreader files? If so, then could someone with some free time on their hands remove all the items from the Wikipedia:Spoken articles list? I don't know the exact number of files there are but it seems like there's quite a few. TheAmazingRaspberry (talk) 00:56, 19 March 2024 (UTC)

Done. Opencooper (talk) 02:29, 19 March 2024 (UTC)

Template guidelines contradict MOS:TALKORDER

Wikipedia:WikiProject Spoken Wikipedia/Template guidelines state that At the top of the article's talk page, add {{WikiProject Spoken Wikipedia}}, emphasis original. This contradicts MOS:TALKORDER, which puts the WP banners below ten other things. Thus, I think this should be reworded to say On the article's talk page, add {{WikiProject Spoken Wikipedia}}, following our guidelines on the order of talk page banners. Queen of Hearts ^she/_they^talk/_stalk 23:18, 17 March 2024 (UTC)