Microsoft AI records 5k audiobooks for Project Gutenberg | Hacker News Reader

Microsoft AI records 5k audiobooks for Project Gutenberg

thenewstack.io

42 pointsbiomcgary17 comments

I'm a big fan of both audible and project gutenberg.

A quick listen of a couple of these recordings reveals some of the worst robotic renderings possible.

Sorry, team, these are awful.

So slightly better than RC Bray? The bar for audio books is *extremely* low. I switched from being a prolific reader to a prolific listener (mostly sci-fi and fantasy) as my family duties had me hauling kids around to activities for hours a day multiple times a week. It's amazing to me how many more novels I just can't get into due to the narrator. I greatly look forward to the day where I can choose my own AI narrator for audio books. With a few notable exceptions, AI generated narration would be a significant improvement.

I don't listen to novels, and my experience isn't the same as yours. I have listened to hundreds of human narrated books and only had poor experiences once or twice due to the narration.

> So slightly better than RC Bray?

Shots fired lol. I don’t think he is that bad, though he does narrate some absolutely trash books (looking at you Helldivers).

I greatly respect and appreciate a good narrator but I think I’m with you, I have some really bad examples of narrators for books I like that I’d gladly replace if given the option. Even better if I can train the AI on narrators I really like. Heck, I’d pay some kind of licensing fee to use their “voice likeness” if that was a thing.

Whoever sells audiobooks….. make a big deal of using human readers.

Good human readers are much, much better than any AI (and much much better than non professional readers).

I buy a lot of audiobooks and it really really matters who they narrator is.

If I found out an audiobook is AI read then I’ll not buy it..,,, not now nor in the future when TTS improves. I want to support good narrators.

Project Gutenberg is making a huge mistake going this hard with AI. They should double down on people and find a way to get more people doing better reads.

All this AI and TTS audiobooks just makes me steer clear of Gutenberg because it’s got to be a huge pile of low quality trash now.

Support human narrators. Say no to AI.

This sounds just like the “use native not electron” arguments to me. Yes, native (human) is better but if the app wouldn’t exist at all without electron (ai), I’ll pick electron.

If Gutenberg gets human narrators for books then that’s great, for the rest using AI doesn’t seem like a bad idea.

Heck, there are 3 books in a series I love (Honor Harrington, first 2 and especially the last one in the Shadow of Saganami series) that have a different narrator from all the mainline series and they bad (I cannot stress enough how bad the narrator for book 4 in the spinoff is). I’d love nothing more than to train an AI on the other books and generate new narrations for those 3 books. Heck I’d pay Allison Johnson $100+ per book for her to re-narrate them if that was an option but I seriously doubt it is with copyright BS and trying to crowdfund something like that. So in the absence of getting what I’d like, and what I’d pay for, I’d gladly turn to “AI”.

Couldn't agree more. Let's also not forget that many people can't actually read for various reasons, such as visual , physical, learning disabilities. Free AI narration is a huge thing for all these people, even if the quality is not perfect. Anything that encourages people to consume more books and build empathy with other humans by experiencing different viewpoints has to be a good thing, right?

Not sure I agree. I listened to a handful of the books that MS encoded. I would not be able to listen to a full book like that. It's just not quite right - the timing is fixed and while the words are correctly pronounced and each have the right accent, there's a lack of natural emphasis on certain words within each sentence that makes if feel cold and alien. I admit this is miles ahead of any previous text-to-speech I've heard, and would be good for reading, say, a blog post or article, but not a full book. At least not fiction. I read fiction for enjoyment and agree that the narrator makes the difference. I've stopped listening to several audiobooks because the narrator didn't cut it for me.

I'll watch this space and reevaluate, but I wouldn't listen to any of this batch.

Thanks for posting this. I wish they had a curated list for the more popular titles. Scanning through the collection is like a treasure hunt. The voices are a little bit robotic, but it is really remarkable how good the technology is getting. We are really starting to cross the uncanny valley here.

Other discussion: https://news.ycombinator.com/item?id=37637812

Link to audiobooks: https://marhamilresearch4.blob.core.windows.net/gutenberg-pu...

Paper: https://arxiv.org/abs/2309.03926

This feels like the equivalent of Amazon spammers creating dozens of junk books for sale in Amazon using chatgpt. The quality of these audiobooks is so poor.

They're free and don't sound horrible, just not award winning.

https://ia801602.us.archive.org/21/items/synapseml_gutenberg...

That's awful.

I think this is unlistenable - it seems to stay on upward intonation as if a question is being asked, for what must be entire paragraphs

Yeah, they are pretty bad.

But if you had said they'd be this good by now just three years ago, I'd have considered you suffering from optimism to the point of delusion.

The nice thing about AI's scalability is that in another 3 years they can run all these books through a better version of voice synthesis that's going to be quite better at inflection than what we have now.

So while a middling win for those with accessibility issues who wanted to hear obscure out of print books due to the questionable quality over a long listen, this is just a step on a longer journey to a product that's much better.

> Microsoft AI records 5k audiobooks for Project Gutenberg

The finally found an use for Narator.

Yes, this happened a while back. I already archived the whole set. It is about 1.25 TB.

Reply on news.ycombinator.com