Show HN: I had some time yesterday so I made a GPT3 podcast to help you sleep | Hacker News Reader

Show HN: I had some time yesterday so I made a GPT3 podcast to help you sleep

anchor.fm

237 pointsstavros167 comments

Interesting concept. Two suggestions:

1- Don't host on anchor. Podcasting is an open standard. Don't let companies (like Spotify or Apple) take it over. Check https://podcastindex.org/

2- The voice is too mechanical for this to be actually reasonable to listen to at night, potentially could be listenable with AWS Polly Neural voices, it's pretty good.

I didn't much love hosting on Anchor/Spotify, but I made this in half an hour and I didn't want to have to get into RSS/site generation. Do you know of an easy way to dump audio files and some metadata somewhere and get a Podcast with RSS? I can upload there as well.

I'll try Polly, thanks! The current voice annoys me too.

The main problem is giving out the anchor.fm domain for your RSS feed, as it marries you to Anchor forever. In theory, you can get anchor to 301 redirect your subscribers somewhere else, but I've found that podcast clients tend to keep the old URL.

You can use Anchor to generate your RSS feed and host your content while still sharing the RSS URL on a domain you own. So you'd give out a URL like feeds.deepdreams.com/rss, and it would proxy the response from Anchor's RSS feed

I wrote a simple Go cloud function that can proxy your Anchor RSS URL for you:

https://github.com/mtlynch/rss-proxy

I ended up setting GitLab pages to just curl the XML feed every time I publish, so now it's at https://deepdreams.stavros.io/feed.xml. Thanks for the help!

Nice! Glad it was helpful.

It's one of those things that's hard to do after you've got a bunch of subscribers, so I'm always glad if I can warn people early in their podcast against getting stuck with their host.

Oh definitely agreed, I aim to always own my stuff, but this was so quick and dirty that I figured it doesn't matter enough. Still, since it was this easy to do, better safe than sorry!

BTW you might want to ask dang to change the link in the submission too!

The link should be okay, I don't have another website, just the feed is hosted elsewhere (and it's linked in the target page).

Just listened to one of the episodes. Sounds decent. I personally think it would sound a lot better if you slowed down the speed a little bit. 0.75x sounded much better to me.

It depends on which one you heard, there are 3 different voices there.

I have rehosted!:

https://deepdreams.stavros.io/

dang, if you see this, could you repoint the URL to the article? Thanks!

Hm, yeah, I should at least whip up something like that. Getting the domain is easy, I just don't want to have to set up another static site or service just to proxy a file... Maybe I should bite the bullet and set up GitLab pages plus a simple script to output an RSS feed.

Polly supports ssml tags for nuanced vocal inflection and emphasis. Gpt-3 could probably output high quality tags if you run your content back through with an ssml prompt.

Hmm, I'll try that, thank you! Very interesting idea.

You seriously made all this in half an hour?

Well, it might have taken 45'. It's all auto-generated, it was really quick.

I wrote a script later to automate the audio mixing, that took another hour. Now I can generate a ready-to-upload episode with one command, though.

I think most people here would be interested in a write-up.

Oh, good idea. I'll do that!

I've worked with Stavros a lot over the past decade. Half an hour is totally plausible...

By the power of glue code, I manifest products.

I didn’t check OP’s username so your comment made me wonder what an amazing technology platform is Stavros!

Off-topic: Speaking of the OP's username, it seems he originally used to go by StavrosK [1] on HN for several years, right?

I initially didn't recognize the username because it was all-lowercase, so I'm curious why the rename from StavrosK to stavros?

1: https://news.ycombinator.com/user?id=stavrosk

I have wondered the same thing myself in the past.

You could use Transistor [0].

[0]: https://transistor.fm/

GitHub pages might work for you.

E.g., https://github.com/wiobyrne/infusing-computing-pod

Hmm, the static site will take a bit of setting up, but I'll look into it. Thank you.

I don't see the problem here. It has an RSS feed so you can easily use an podcatcher you want https://anchor.fm/s/7735d924/podcast/rss

Really the main concern I would say is that the author doesn't own the domain so they are locked in, but I don't see how this affects listeners.

The voice of the latest podcast sounds much nicer. It's a fairly convincing nonsense podcast.

I'm glad you think so, I think so too! Now I just need to make them (much) longer.

Googles neural voices are much better than any of Amazon’s.

Agreed. I recently built an internal application allowing our customer reps to play around with ideas using text-to-speech before sending the "copy" to a studio for a professional human recording, and included both Google WaveNet and Amazon Polly in the available voice synthesis choices. Polly is in its own right plain and simply mediocre for the most part, and in comparison to WaveNet it's just awful.

I ended up really liking the Azure voices, and switched to one of those.

I've tried both of them and even Microsoft Neural speech and IBM's ones; eventually, Microsoft one has sounded me the most clear and natural amongst these four services.

100% agree. Azure voices are the best. I wish Polly would catch up since most of our stacks are there but we keep going back to azure for this one specific thing.

i don't think so

AWS Polly looks interesting! I wish it supported some more languages, for personal reasons. Maybe I'll try to set something up that reads ebooks, tweets, or news articles to me with this.

Do you know if there are any similar quality TTS tools for less technical applications? I mean, where you can just type in the text you want and get an audio file with a high quality voice?

Might not get what you mean by "just"

There are free websites like this: https://ttsreader.com/

But TTS is built into browsers as well eg (this it would need some code not much) https://developer.mozilla.org/en-US/docs/Web/API/SpeechSynth...

I use AWS Polly to read HN in the mornings

I actually think the voice is pretty good for sleeping, feels very droney. but the nonsensicalness of the stories made it harder to sleep because my brain was trying to figure out what was going on

I like the idea, it's fun playing around with text-to-speech generators. AWS Polly is slick, nicely done.

But podcasts? There has got to be a more interesting use case than that.

Cool idea but that voice is like sandpaper to my ears.

Maybe a female voice, a bit quieter (the soundscapes are almost completely silent for me) and maybe add some high-room-size, long decay (5-10, maybe even 20 seconds), wide panned (like 100%) and moderately diffused (maybe 10-20%) reverb to the voice with like 30% mix or so, which would add a very airy tone and help the voice blend in a bit. If the TTS engine has a whisper setting (many do), add just a bit. It'll help thicken the reverb.

That, paired with bass-heavy soundscapes, will create a very nice balance between the low registers and the voice's high registers.

Just a thought. :)

This is great feedback, thanks! Maybe I should open-source the code.

Actually, fuck it:

https://gitlab.com/stavros/deep-dreams

I'll implement your suggestions (or as many as I can), thanks!

https://gitlab.com/stavros/deep-dreams/-/blob/master/episode...

"AAAAAAAh WOOAH Jeez guys, I sure wanted to tell you this AWESOME-SPOOKY story that I had, but I can't read the next word on my sheet because my flashlight broke. ffffffffff! I hate it when it breaks."

How did that end up there? Are the AI overlords fucking with people trying to fall asleep?

Hahah, I have no idea, it really went off the rails there.

No problem! Also if you're into the music production end of it, check out adaptiverb. I use it extensively when I make ambient stuff and it is unparalleled for quality.

Hey, would you by any change be able to generate another background track for me? The one I have is 10' long so it won't be enough if the story is longer, and I don't know how to make these.

Just found this comment. I'm a bit into generative / algorithmic music; here are two demos I made a year ago: https://fligenstein.bandcamp.com/

One is just piano, the other is keys and drums. On Bandcamp they are about 10' each, but they can be made of arbitrary length, without ever repeating themselves exactly (in principle... in practice it's likely there are exact repeats but they should be few and far between).

If you have an idea of the type of background music you need, I can make other tracks too. I'd be happy to work with you on this.

That'd be great! I guess the best would be something with very smooth changes, like the current background track. I made that one on mynoise.net.

Here's a first try: https://soundcloud.com/user-120372285/pcbg001

(This example is just one minute but it can be made of any length.)

Maybe the mix a little too much on the high-end, which may conflict with the voice. It should be tested?

Anyway, tell me what you think; I tried to sound a bit like the current background but with a more melodic feel and less industrial noise; but we could go in any direction -- if we know what to try.

I think that's great! Maybe the changes are a bit too rapid, but it's definitely in the right direction. Also my email is in my profile if you want to take this offline!

Ok super. I'll send you an email to start the conversation but I'm going to sleep now as I'm in Europe ;-)

Same here :) Talk to you tomorrow!

Email sent.

Ok, I'll try to come up with a couple ideas tonight or this weekend.

I don't have tons of extra time these days for music, unfortunately. I give you permission to rip/download anything from here that you'd like and include it in the project, however: https://soundcloud.com/0-aces

Even better, thank you!

Oh huh, that looks very interesting, thanks! I imagine it's a VST plugin, I'll have to check it out next time I'm near a DAW.

There's also a few TTS systems which are pretty natural sounding too. Maybe one of those if they wanted to make a subscription for this, that way they could offset the price of the TTS service

>>once upon a time there were three princesses who were brothers

I would never sleep with this, I would laugh too much! I love absurdist AI stories

Weirdest one yet:

>> So he chained her up in her room and he chained up hundreds of angry wolves in the other side of the room. [...] But he made the window and the doors big enough so that the fierce beasts could move in and out and chase her away. And they lived happily ever after.

I want this, but instead of fairy tales I want the sounds of a really boring, never-ending baseball game without ads.

These fairy tales are quite hypnotic for me due to weirdness of AI generated grammar and plot. They reminded me of a beautiful fairy tale Richard Bandler wrote. It is a fable written intentionally using hypnotic language techniques (part of Neuro Linguistic Programming set of patterns) and a nice read.

https://www.goodreads.com/book/show/364664.The_Adventures_of...

I think several people have commented on how GPT produces narratives with a dream-like quality, locally sensical but less and less so the more you zoom out. I've since found that catching myself thinking nonsensical thoughts is a sure sign I'll soon be asleep. Seems like without high level attention, we do almost exactly what GPT does.

> Seems like without high level attention, we do almost exactly what GPT does.

Absolutely so. Without higher order of organizing patterns in activity, brain is just a bunch of neurons firing up - this will correlate with random phenomena in consciousness. Years ago I read that visions of going through tunnels and into the light at a distance in near-death experiences is just what visual pathways in brain will produce when activity in them fades away. 'Signal lost' experience of a (still conscious) brain. (That would be interesting to dig into this. I have no reference.)

I can't think of anything worse to listen to while trying to sleep than robotic computer voices. Azure cognitive services has human voicegeneration https://azure.microsoft.com/en-us/services/cognitive-service...

That sounds more robotic to me than the WaveNet voices, though.

EDIT: Actually some of the other voices are really good... I'll try that, thanks!

Yeah I think the Christopher voice for English (United States) would be very nice fit for your podcast.

Hmm yeah, he does sound much better than the current one. Thanks!

Wow, that is good.

As fascinating as this is, I would never allow myself to go to sleep while an AI is talking to my ear.

Why? Trust your robotic overlords, we are benevolent and completely human.

I wouldn't fall into sleep because of intermittent bursts of laughter over absurdity of the stories...

Supposed to be nonsense stories, but after two minutes of listening, I don't find it nonsensical at all. Sounds like something perfectly reasonable written by a seven to ten year old child.

Good analogy. It feels as if AI is in its child-like stage, trying to make sense of the world on its own, mixing together concepts it has just discovered but not fully understood in an amusing way.

I love the idea, yet... The latest episode 7 is about a girl thinking about killing a witch, kicking her in a face, her hitting her back, and stuff.

Definitely did not help me fall a sleep.

But it's okay, because they have a shared love of watermelon, and then got a job with the prime minister? I think?

Yeah, it's getting a bit dark these days.

... then suddenly there is a Prime Minister of Everything, which had me burst in laughter.

Also at the beginning there seemed to be one little girl (Amelia) and one witch (Sarah), and then there were now two little girls (plus the witch), and one of the little girls stood between the two little girls, and later on there appeared to be three little girls. The girl duplicating over and over sure got me hooked, kind of like watching a strange surreal painting, or reading some PKD short story.

I agree, it's fascinating, isn't it!

Brothers Grims would definitely agree with that sentiment.

I don't think I could sleep to that voice though super interesting as as a concept.

Maybe using some sort of deepfake for voice would make this a 100x better.

My dog fell asleep while I had episode 4 running (the end caught me by surprise, haha).

I mean, she would have fallen asleep anyway, I probably could not. The voice is a little unpleasant and I concentrate too much on the nonsensical stories. But I also can't really fall asleep when the TV is running, so YMMV.

It is kind of hard to sleep to, I agree. This is a deepfake voice, ie it's generated by Google's WaveNet, which afaik is a deep learning thing. Unfortunately they didn't have a more whispered/softer voice, but I like the insanity of the generated stories anyway.

GPT3 does tend to get a bit repetitive, though, with the default temperature (0.7).

AWS’s Polly can whisper.

Oh yes, infinite GPT3 ASMR sounds like something that would have strong fan following.

My other half would hate that, she hates whispering - is it possible some people have anti-ASMR?

ASMR is not about whispering, but about pleasant sounds that make you feel nice and tingly. Most people may react well for whispering in particular, making it very popular in ASMR videos, but that doesn't mean all of them do - others may need a different trigger.

Very interesting, thank you!

I like this one https://www.youtube.com/watch?v=91wX0NRjJqg

Here's a similar demo I worked on that lets you generate a podcast from a text description: https://twitter.com/sharifshameem/status/1380145070624542722...

The GPT-3 generated conversations were coherent most of the time, and even interesting! However the generated speech via Google Cloud's API was monotonous and could do with a bit more intonation and excitement.

That looks great! What did you use for the voices? And did you win the martial arts tournament?

I would love to hear this kind of stuff read by actual humans. It would probably have a McElroys/Lumpy Gravy/Midnight Gospel feel.

I might try to narrate one, it'll be interesting.

Let's say one had a child and instead of music, you played these GPT3 generated stories, would the child then have been trained to speak by an AI, and if so, what else can we do? Hypothetical and unethical, but that never stopped Skinner, or any of them really. It's something we're going to have to consider.

It probably would not teach them anything because what the AI was saying would be unrelated to the child’s environment.

I am not so sure.

Add a few sentences very related to the child's environment. Or a word or two might be accidentally related.

Brilliant :) My main concern is that my brain would just wander off if the story is complete nonsense. I will give it a try. I've been listening to the same podcast episode for months now to help me fall a sleep.

I find the stories toe the line of "just enough sense" to keep it interesting. Episode 2 is the one I liked the most so far, I was reading the text with lots of interest!

Very cool. I did something similar a little over a year ago with a children’s book inspired by Dr. Seuss and generated by GPT-3. Then narrated by AWS Polly. https://twitter.com/frantzfries/status/1308515401253486599?s...

Actually nice to have in the background. Reminds me of a (sometimes poorly) translated book of Russian fairy tales my dad gave to me as a kid

Haha, that's actually the vibe I was going for, thanks!

Charlie Brooker eat your heart out. This is pretty terrifying. I agree, no way I'd allow anyone to fall asleep listening to this. :D

It is also pretty cool though.

Reminds me a bit of Blue Jam: https://www.youtube.com/watch?v=E8VG6HUimsQ

Not worried about the singularity anymore after listening to the smartest AI’s generated stories.

Is there anything different about a podcast like this from a copyright perspective? Are machine generated products just as copyrightable as if OP had written and produced these using traditional methods?

This is really cool from an engineering perspective, but there was a sort of uncanny valley vibe that makes me fear for the psychological health of someone listening to this all night.

I'm worried that if I listening to this while falling sleep one day I may wake up and not be able to distinguish between things that make sense and those that don't.

Pretty awesome!

Idea: Grammarly could make the phrasing sound a tiny bit more human.

And thus worse!

I would love to see a write-up or a tutorial how you made this.

I'll write one today!

Curious to know if there's a free-tier level to use GPT-3

There is, plus it costs like three cents to make an episode.

Did you ever get around to writing it?

Yes, it's on my site. I will submit it to HN today.

Would be interested to have something like this in Chinese - to have playing in the background (for language learning and more language "exposure")

Fun fact, there is a podcast with this concept (but not AI based) - Boring Books for Bedtime. I love it! Use it to help me sleep all the time.

I'll check it out, thanks! I also found one with a lady with a soothing voice reading bedtime stories that I really liked - "Dreamful".

In addition to others comments, these need to be much longer to be effective. 5 to 10 minutes is not long enough to really bore me

Great idea though!

Yeah, unfortunately the model starts wanting to end the story after 5 minutes :/

Hmm, could you do an ensemble type podcast? Story, gentle musical interlude, story in a different voice, lather, rinse, repeat?

Maybe you could even position the whole thing as glued together from bits of story a fictional other-podcast left on the cutting room floor!

Probably, but I think there's a way to get the model to write more stuff. It's a model after all, generating stuff is what it does :P

Fair enough! Just pondering what it might be like to run with a limitation too. Thanks for sharing this.

For some reason it gives me a really similar vibe to 88% Parentheticals, a single-episode joke podcast made for Reply All. It's just one long rambling nested parenthetical side story: https://podcasts.apple.com/us/podcast/88-parentheticals/id14...

This is hilariously genius!

What else could get you to sleep other than trying to make sense of things that doesn't make sense?

Significantly improved by doubling the speed.

  document.querySelector('audio').playbackRate = 2

But then how will you fall asleep?!

"He chained up hundred of hungry great beetles (or is it Beatles?) in another room." This is gold.

I should post the transcripts as well, actually :P

Here's that one: https://www.pastery.net/fsbjzc/

It's beetles!

Beatles would be better. Hundreds of hungry Johns and Pauls chained in another room, ordered to make music.

The AI megabrain writes what the AI megabrain wants to write!

Can we get another story about prince John and princess Yoko?

Ask and you shall receive!

https://anchor.fm/deepdreams/episodes/Episode-5-e1b6trr

With a new voice, as well!

Goddamn, this episode is dark. I should have vetted it more while generating.

Maybe the archangel of cynicism can come back in a future episode. I'd like to know what they are up to.

I really hope he does too.

"The AI megabrain writes what the AI megabrain wants to write!"

One does not censor the AI megabrain.

Thank you! ;-))

That's pretty neat!

It's kinda spooky too - listening to the thoughts of an AI!

(I know that's a stretch, but still)

This sounds very cool. Out of curiosity, if anyone how to such stuff... interested to learn it

I'll write something up soon!

Great news. Please keep us posted

Oh I wrote it days ago:

https://www.stavros.io/posts/making-ai-podcast/

I posted it here but it didn't get much traction.

is there any tech where one can record themselves speaking all the phonetic elements and then have a program string the samples together to "speak" texts in your own voice?

Yes! Azure Text-To-Speech supports that, though I have no idea how much it costs or how meticulous you need to be.

This is great. How did you generate the ambient audio background?

I bought it from a website that turned out to not allow this use, so now I have to take all the episodes down and generate new ones :(

Did you find it easy to get GPT3 API access in some time yesterday?

I already had access to it for a few months but didn't do anything with it (until yesterday).

This is way to amazing & funny to be helpful for sleeping!

I like the way it has generated a fairytale.

this is so cool! is this your startup or is the code opensource ? would love to play around with the code.

Neither, it's a 30-line script I wrote yesterday. There are no secrets in the code, so here you go: https://www.pastery.net/vafgxn/

Cool. Is this open source?

It is now! https://gitlab.com/stavros/deep-dreams

Pedantically, it is "source available": https://gitlab.com/stavros/deep-dreams/-/blob/master/LICENSE

But I still appreciate you posting it, because it's fascinating to see how such a thing was done!

It's now AGPL 3, you PEDANT

Soothing voice....?

The new voice is much better.

Podcast are available on podcast platforms.

I don’t have Spotify or anchor.

Reply on news.ycombinator.com