AI – Nitin Khanna

Last night I came across a ListenLater.net which has an interesting value proposition – send them text or the link to an article and they’ll convert it into a podcast using AI TTS. The podcast link will be public so you can use it in your favorite podcast player, which is such a nice touch! The voice they used seemed familiar but I couldn’t immediately place it.

Digging into their help pages or pricing didn’t give me a lot of details about how they’re generating the audio. They just keep claiming it’s “advanced AI Text-to-Speech”.

Their EULA says you can’t use the audio for commercial use. It has to be personal use only. This is partly because they acknowledge that they claim no ownership to the content you send to them and so if you use it commercially, they don’t want to be held liable for that.

But that voice…

In a spectacular feat of google-fu, I typed in “What TTS is listenlater.net using?”

I learnt that there is a similar service called Listenlater.fm which uses a horrible non-neural TTS (feedback from HackerNews) which is unbearable. Also, though the site is up, the audio samples are not available, which tells me that maybe that service isn’t doing so well. But also, their pricing model is funky – 5 free articles per month and if you want more, $36/year for unlimited.

Listenlater.net instead uses a more AI-aligned pricing of $0.03 per 1000 tokens (about 750 words according to them). This is a clear indication that they’re using a third party service without telling us which.

I then came across a service called listnr.ai which… takes text and gives you a podcast. You can also use their output for YouTube videos, TikToks, Reels, Shorts, Gaming, Social Media, and audiobooks. (Also, they’ve done a nice job of comparing their service to others in the same space. Thanks for doing the market-research for me, folks!)

Except… their terms say you can only use the content you download from their site for “personal, non-commercial use”. So… their own sales are violating their own TOS?

But the service is in India, so I guess they can ignore these rules.

But what’s the point of finding listnr.ai if I can’t validate that they have the same voice as Listenlater.net? It has to match! I listened to 50 voices and the absolutely last one, called “Shimmer” was a match!

Ok, but where is Shimmer coming from? I don’t trust listnr.ai to have built their own AI TTS just like I don’t trust listenlater.net to have.

Back to the Google-board! “Shimmer tts voice”

The first few results are some shitty site called 101soundboards.com and then one from a listnr.ai competitor called FakeYou. Then, below the Google fold of “People also ask”, we get the result we’re looking for. Mirroring the last 6 voices that listnr.ai supports are –

OpenAI’s alloy , echo , fable , onyx , nova , and shimmer.

Ah. There it is. I listened to a sample and sure enough, it matches exactly what listnr.ai is selling and very, very close to listenlater.net’s primary voice. So both these services are basically built on top of OpenAI and they just don’t want to talk about it. Why?

Well, OpenAI’s TTS documentation page says “Please note that our usage policies require you to provide a clear disclosure to end users that the TTS voice they are hearing is AI-generated and not a human voice.”

So while they’re very happy telling you that you’ll be listening to AI TTS, if you use the audio commercially and OpenAI comes after you, these companies want to protect their businesses. Nothing wrong with that.

Also, nothing wrong with reselling OpenAI’s service either. The service is API based. So normal users can’t use it. Building a website, a service, a podcast hosting setup, and supporting all this takes Engineering and Business hours and is well worth the added cost that these services might be pushing to their users. In the case of listenlater.net, it seems that’s not true either. OpenAI charges $30 per 1 million characters for their HD voices and $15 for non-HD. The difference is quality vs speed. Listenlater.net charges the same – $0.03 per 1 thousand characters. So if they’re not using the HD TTS, they’re pocketing half the money. Or they’re not and you’re getting a service that’s running out of love.

Listnr.ai’s pricing is a little more FU – it starts at 4000 words per month for $5 per month. But considering they are adding a lot more bells and whistles to their services – unlimited downloads and audio embeds, 25GB storage, 1000 voices (I didn’t bother finding out where they’re getting their other voices. Most seem to be coming from ElevenLabs, including Santa Clause. Exercise left to the reader), it might be worth it to someone out there.

Anyways, good dive.

I recently purchased an Intel N100 based mini PC with the idea of turning it into a local webserver hosting many services like RSS feed reader, pi-hole, an open source memos app, and a few other smaller open source projects. I currently host all of these on an ailing Macbook Pro which has a tendency to go into “darkwake”, i.e. it wakes up when I send it a Wake-on-LAN command, and then within 15 seconds it goes back to sleep. Rather a lethargic fellow, that.

Initially wanting to install ubuntu on this new machine, I opted for debian because I’ve recently had some interaction with the OS at work and I’ve found it to be very stable and light. I’m sure someone might disagree and I’ll gladly experiment with other Linux OS somewhere down the line.

While I was installing debian, it asked me what the domain should be and I answered “local” because I want to access the server on my local network with the completely innovative domain name “server.local”. My other machines are “laptop.local” and “macserver.local”.

The latter two are Macbooks, so they automatically advertise the .local domain on the network using Apple Bonjour. But the debian wasn’t doing so. I was googling around but didn’t even know what to ask. Some stackoverflow answers spoke of how Ubuntu automatically advertises the “.local” while debian does not.

Before I rued my fate and having to wipe out my just setup server, I decided to ask the AI powers how I can solve this problem. My go-to is Bing Chat, since it has internet access. I asked the

Does debian 12 advertise itself on the lan with a .local domain name like Ubuntu and MacOS do?

Bing Chat looked around and regurgitated an answer I’d already found on Stackoverflow – that while it is possible to do so, the SO answer author decided to leave out the vital detail of what the heck this service is called on the Linux side of things (Apple Bonjour is a damn well known name in tech circles). The alternatives that the SO answer mentioned and Bing Chat vomited were to setup a local DNS server or to use /etc/hosts. The latter option is NOT available on iOS devices, which are painfully inadequate in terms of actually completely owning your device.

Before I gave up, I went to ChatGPT and pasted the question above exactly as is. I wasn’t really expecting a different answer, but I sure got it.

According to ChatGPT, as of its January 2022 cutoff date, Debian 12 hadn’t been released. But, the technology I’m talking about is called Avahi. Ubuntu and MacOS have it preinstalled and I can check if Avahi is installed using the following command –

dpkg -l | grep avahi

My debian install was of the netinstall flavor, which means it installed all the basic packages it thought were relevant and everything else was left to this poor user to figure out. I googled and found a method to install avahi-daemon on debian and the tutorial even mentioned that after installing it, I basically have to do nothing.

Lo and behold! One quick install step later, I can now access “server.local” on my LAN. Nice!

I don’t often do this, but I dropped the following feedback to the ChatGPT team regarding this excellent answer their LLM provided to me –

Though it doesn’t know that Debian 12 has been released after its cutoff date, ChatGPT was nevertheless point me in the right direction because the underlying technology – avahi – has been around for a while and is clearly the answer to the question I was looking for. In contract, Bing Chat was not able to come to the same answer even though I asked it the same question.

Spelling mistake and all.

The title of this post was created by ChatGPT after I fed it the entire post above and asked it what to name the post. I gave it the title “ChatGPT still wins over Bing Chat”, but that felt too sensational.

Cheers.